modelId
stringlengths 5
139
| author
stringlengths 2
42
| last_modified
timestamp[us, tz=UTC]date 2020-02-15 11:33:14
2025-08-30 12:27:52
| downloads
int64 0
223M
| likes
int64 0
11.7k
| library_name
stringclasses 528
values | tags
listlengths 1
4.05k
| pipeline_tag
stringclasses 55
values | createdAt
timestamp[us, tz=UTC]date 2022-03-02 23:29:04
2025-08-30 12:27:19
| card
stringlengths 11
1.01M
|
|---|---|---|---|---|---|---|---|---|---|
mlx-community/Qwen3-14B-4bit-AWQ
|
mlx-community
| 2025-08-29T20:42:39Z | 1,142 | 3 |
mlx
|
[
"mlx",
"safetensors",
"qwen3",
"unsloth",
"text-generation",
"conversational",
"base_model:unsloth/Qwen3-14B",
"base_model:quantized:unsloth/Qwen3-14B",
"license:apache-2.0",
"4-bit",
"region:us"
] |
text-generation
| 2025-05-06T15:22:57Z |
---
tags:
- unsloth
- mlx
base_model: unsloth/Qwen3-14B
license: apache-2.0
pipeline_tag: text-generation
library_name: mlx
---
# mlx-community/Qwen3-14B-4bit-AWQ
This model [mlx-community/Qwen3-14B-4bit-AWQ](https://huggingface.co/mlx-community/Qwen3-14B-4bit-AWQ) was
converted to MLX format from [unsloth/Qwen3-14B](https://huggingface.co/unsloth/Qwen3-14B)
using mlx-lm version **0.26.3**.
## Use with mlx
```bash
pip install mlx-lm
```
```python
from mlx_lm import load, generate
model, tokenizer = load("mlx-community/Qwen3-14B-4bit-AWQ")
prompt = "hello"
if tokenizer.chat_template is not None:
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
messages, add_generation_prompt=True
)
response = generate(model, tokenizer, prompt=prompt, verbose=True)
```
|
8septiadi8/blockassist-bc-curious_lightfooted_mouse_1756500004
|
8septiadi8
| 2025-08-29T20:41:24Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"curious lightfooted mouse",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:41:19Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- curious lightfooted mouse
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
Fentible/Cydoniathulhu-24B-v1-GGUF
|
Fentible
| 2025-08-29T20:40:47Z | 233 | 0 |
transformers
|
[
"transformers",
"gguf",
"mergekit",
"merge",
"text-generation",
"en",
"base_model:Darkhn/M3.2-24B-Animus-V7.1",
"base_model:merge:Darkhn/M3.2-24B-Animus-V7.1",
"base_model:Doctor-Shotgun/MS3.2-24B-Magnum-Diamond",
"base_model:merge:Doctor-Shotgun/MS3.2-24B-Magnum-Diamond",
"base_model:Gryphe/Codex-24B-Small-3.2",
"base_model:merge:Gryphe/Codex-24B-Small-3.2",
"base_model:ReadyArt/MS3.2-The-Omega-Directive-24B-Unslop-v2.1",
"base_model:merge:ReadyArt/MS3.2-The-Omega-Directive-24B-Unslop-v2.1",
"base_model:TheDrummer/Cydonia-24B-v4.1",
"base_model:merge:TheDrummer/Cydonia-24B-v4.1",
"base_model:zerofata/MS3.2-PaintedFantasy-v2-24B",
"base_model:merge:zerofata/MS3.2-PaintedFantasy-v2-24B",
"license:apache-2.0",
"endpoints_compatible",
"region:us",
"conversational"
] |
text-generation
| 2025-08-22T09:13:24Z |
---
base_model:
- Darkhn/M3.2-24B-Animus-V7.1
- Doctor-Shotgun/MS3.2-24B-Magnum-Diamond
- Gryphe/Codex-24B-Small-3.2
- ReadyArt/MS3.2-The-Omega-Directive-24B-Unslop-v2.1
- TheDrummer/Cydonia-24B-v4.1
- zerofata/MS3.2-PaintedFantasy-v2-24B
emoji: 🐙
language:
- en
library_name: transformers
license: apache-2.0
tags:
- mergekit
- merge
pipeline_tag: text-generation
---
# 🐙 Cydoniathulhu 24B v1 GGUF
Experimental side branch of the Cthulhu series with an emphasis on Cydonia.
```
base_model: TheDrummer/Cydonia-24B-v4.1
merge_method: dare_ties
architecture: MistralForCausalLM
dtype: bfloat16
models:
- model: Darkhn/M3.2-24B-Animus-V7.1
parameters:
density: 0.5
weight: 0.16
- model: Doctor-Shotgun/MS3.2-24B-Magnum-Diamond
parameters:
density: 0.5
weight: 0.16
- model: Gryphe/Codex-24B-Small-3.2
parameters:
density: 0.5
weight: 0.16
- model: ReadyArt/MS3.2-The-Omega-Directive-24B-Unslop-v2.1
parameters:
density: 0.5
weight: 0.16
- model: TheDrummer/Cydonia-24B-v4.1
parameters:
density: 0.5
weight: 0.2
- model: zerofata/MS3.2-PaintedFantasy-v2-24B
parameters:
density: 0.5
weight: 0.16
tokenizer:
source: union
chat_template: auto
```
|
KoKoDanio/cyber_llama_32
|
KoKoDanio
| 2025-08-29T20:39:16Z | 0 | 0 | null |
[
"safetensors",
"unsloth",
"text-generation",
"en",
"dataset:Rowden/CybersecurityQAA",
"base_model:meta-llama/Llama-3.1-8B-Instruct",
"base_model:finetune:meta-llama/Llama-3.1-8B-Instruct",
"license:mit",
"region:us"
] |
text-generation
| 2025-08-29T19:58:05Z |
---
license: mit
tags:
- unsloth
datasets:
- Rowden/CybersecurityQAA
language:
- en
base_model:
- meta-llama/Llama-3.1-8B-Instruct
pipeline_tag: text-generation
---
|
liukevin666/blockassist-bc-yawning_striped_cassowary_1756499837
|
liukevin666
| 2025-08-29T20:38:25Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"yawning striped cassowary",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:38:13Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- yawning striped cassowary
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
hitrax/blockassist-bc-timid_toothy_meerkat_1756499819
|
hitrax
| 2025-08-29T20:38:14Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"timid toothy meerkat",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:38:09Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- timid toothy meerkat
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
gigamega/elysia
|
gigamega
| 2025-08-29T20:36:58Z | 0 | 0 |
diffusers
|
[
"diffusers",
"text-to-image",
"lora",
"template:diffusion-lora",
"base_model:black-forest-labs/FLUX.1-dev",
"base_model:adapter:black-forest-labs/FLUX.1-dev",
"region:us"
] |
text-to-image
| 2025-08-29T20:35:21Z |
---
tags:
- text-to-image
- lora
- diffusers
- template:diffusion-lora
widget:
- output:
url: images/download copy_crop.png
text: '-'
base_model: black-forest-labs/FLUX.1-dev
instance_prompt: elysia202508
---
# Elysia
<Gallery />
## Trigger words
You should use `elysia202508` to trigger the image generation.
## Download model
[Download](/gigamega/elysia/tree/main) them in the Files & versions tab.
|
bah63843/blockassist-bc-plump_fast_antelope_1756499610
|
bah63843
| 2025-08-29T20:34:36Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"plump fast antelope",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:34:12Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- plump fast antelope
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
huggingtoots/TheDrummer-GLM-Steam-106B-A12B-v1-MLX-4Bit
|
huggingtoots
| 2025-08-29T20:34:19Z | 0 | 0 |
mlx
|
[
"mlx",
"safetensors",
"glm4_moe",
"base_model:TheDrummer/GLM-Steam-106B-A12B-v1",
"base_model:quantized:TheDrummer/GLM-Steam-106B-A12B-v1",
"4-bit",
"region:us"
] | null | 2025-08-29T20:23:28Z |
---
base_model: TheDrummer/GLM-Steam-106B-A12B-v1
tags:
- mlx
---
# huggingtoots/TheDrummer-GLM-Steam-106B-A12B-v1-MLX-4Bit
The Model [huggingtoots/TheDrummer-GLM-Steam-106B-A12B-v1-MLX-4Bit](https://huggingface.co/huggingtoots/TheDrummer-GLM-Steam-106B-A12B-v1-MLX-4Bit) was converted to MLX format from [TheDrummer/GLM-Steam-106B-A12B-v1](https://huggingface.co/TheDrummer/GLM-Steam-106B-A12B-v1) using mlx-lm version **0.26.4**.
🦛 <span style="color:#800080">If you want a free consulting session, </span>[fill out this form](https://forms.gle/xM9gw1urhypC4bWS6) <span style="color:#800080">to get in touch!</span> 🤗
## Use with mlx
```bash
pip install mlx-lm
```
```python
from mlx_lm import load, generate
model, tokenizer = load("huggingtoots/GLM-Steam-106B-A12B-v1-mlx-4Bit")
prompt="hello"
if hasattr(tokenizer, "apply_chat_template") and tokenizer.chat_template is not None:
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
response = generate(model, tokenizer, prompt=prompt, verbose=True)
```
|
elmenbillion/blockassist-bc-beaked_sharp_otter_1756498105
|
elmenbillion
| 2025-08-29T20:33:30Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"beaked sharp otter",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:33:26Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- beaked sharp otter
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
duppbuy/blockassist-bc-scented_scurrying_bobcat_1756498990
|
duppbuy
| 2025-08-29T20:23:44Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"scented scurrying bobcat",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:23:11Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- scented scurrying bobcat
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
jerryzh168/Qwen3-8B-INT4-non-hqq
|
jerryzh168
| 2025-08-29T20:22:46Z | 20 | 0 |
transformers
|
[
"transformers",
"pytorch",
"qwen3",
"text-generation",
"torchao",
"conversational",
"en",
"arxiv:2507.16099",
"base_model:Qwen/Qwen3-8B",
"base_model:quantized:Qwen/Qwen3-8B",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"endpoints_compatible",
"region:us"
] |
text-generation
| 2025-08-19T18:12:30Z |
---
base_model: Qwen/Qwen3-8B
tags:
- transformers
- torchao
- qwen3
license: apache-2.0
language:
- en
---
# INT4 Qwen/Qwen3-8B model
- **Developed by:** jerryzh168
- **License:** apache-2.0
- **Quantized from Model :** Qwen/Qwen3-8B
- **Quantization Method :** INT4
# Inference with vLLM
Install vllm nightly and torchao nightly to get some recent changes:
```
pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
pip install torchao
```
## Serving
Then we can serve with the following command:
```Shell
# Server
export MODEL=jerryzh168/Qwen3-8B-INT4
VLLM_DISABLE_COMPILE_CACHE=1 vllm serve $MODEL --tokenizer $MODEL -O3
```
```Shell
# Client
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "jerryzh168/Qwen3-8B-INT4",
"messages": [
{"role": "user", "content": "Give me a short introduction to large language models."}
],
"temperature": 0.6,
"top_p": 0.95,
"top_k": 20,
"max_tokens": 32768
}'
```
Note: please use `VLLM_DISABLE_COMPILE_CACHE=1` to disable compile cache when running this code, e.g. `VLLM_DISABLE_COMPILE_CACHE=1 python example.py`, since there are some issues with the composability of compile in vLLM and torchao,
this is expected be resolved in pytorch 2.8.
# Inference with Transformers
Install the required packages:
```Shell
pip install git+https://github.com/huggingface/transformers@main
pip install torchao
pip install torch
pip install accelerate
```
Example:
```Py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "jerryzh168/Qwen3-8B-INT4"
# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)
# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
# conduct text completion
generated_ids = model.generate(
**model_inputs,
max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
# parsing thinking content
try:
# rindex finding 151668 (</think>)
index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
index = 0
thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("
")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("
")
print("thinking content:", thinking_content)
print("content:", content)
```
# Quantization Recipe
Install the required packages:
```Shell
pip install git+https://github.com/huggingface/transformers@main
pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126
pip install torch
pip install accelerate
```
Use the following code to get the quantized model:
```Py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig
model_id = "Qwen/Qwen3-8B"
model_to_quantize = "Qwen/Qwen3-8B"
from torchao.quantization import Int4WeightOnlyConfig
quant_config = Int4WeightOnlyConfig(group_size=128, use_hqq=True)
quantization_config = TorchAoConfig(quant_type=quant_config)
quantized_model = AutoModelForCausalLM.from_pretrained(model_to_quantize, device_map="auto", torch_dtype=torch.bfloat16, quantization_config=quantization_config)
tokenizer = AutoTokenizer.from_pretrained(model_id)
# Push to hub
USER_ID = "YOUR_USER_ID"
MODEL_NAME = model_id.split("/")[-1]
save_to = f"{USER_ID}/{MODEL_NAME}-INT4"
quantized_model.push_to_hub(save_to, safe_serialization=False)
tokenizer.push_to_hub(save_to)
# Manual Testing
prompt = "Hey, are you conscious? Can you talk to me?"
messages = [
{
"role": "system",
"content": "",
},
{"role": "user", "content": prompt},
]
templated_prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
print("Prompt:", prompt)
print("Templated prompt:", templated_prompt)
inputs = tokenizer(
templated_prompt,
return_tensors="pt",
).to("cuda")
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128)
output_text = tokenizer.batch_decode(
generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Response:", output_text[0][len(prompt):])
```
Note: to `push_to_hub` you need to run
```Shell
pip install -U "huggingface_hub[cli]"
huggingface-cli login
```
and use a token with write access, from https://huggingface.co/settings/tokens
# Model Quality
We rely on [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) to evaluate the quality of the quantized model. Here we only run on mmlu for sanity check.
| Benchmark | | |
|----------------------------------|----------------|---------------------------|
| | Qwen/Qwen3-8B | jerryzh168/Qwen3-8B-INT4 |
| mmlu | To be filled | To be filled |
<details>
<summary> Reproduce Model Quality Results </summary>
Need to install lm-eval from source:
https://github.com/EleutherAI/lm-evaluation-harness#install
## baseline
```Shell
lm_eval --model hf --model_args pretrained=Qwen/Qwen3-8B --tasks mmlu --device cuda:0 --batch_size 8
```
## INT4
```Shell
export MODEL=jerryzh168/Qwen3-8B-INT4
lm_eval --model hf --model_args pretrained=$MODEL --tasks mmlu --device cuda:0 --batch_size 8
```
</details>
# Peak Memory Usage
## Results
| Benchmark | | |
|------------------|----------------|--------------------------------|
| | Qwen/Qwen3-8B | jerryzh168/Qwen3-8B-INT4 |
| Peak Memory (GB) | To be filled | To be filled (?% reduction) |
<details>
<summary> Reproduce Peak Memory Usage Results </summary>
We can use the following code to get a sense of peak memory usage during inference:
```Py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig
# use "Qwen/Qwen3-8B" or "jerryzh168/Qwen3-8B-INT4"
model_id = "jerryzh168/Qwen3-8B-INT4"
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_id)
torch.cuda.reset_peak_memory_stats()
prompt = "Hey, are you conscious? Can you talk to me?"
messages = [
{
"role": "system",
"content": "",
},
{"role": "user", "content": prompt},
]
templated_prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
print("Prompt:", prompt)
print("Templated prompt:", templated_prompt)
inputs = tokenizer(
templated_prompt,
return_tensors="pt",
).to("cuda")
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128)
output_text = tokenizer.batch_decode(
generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Response:", output_text[0][len(prompt):])
mem = torch.cuda.max_memory_reserved() / 1e9
print(f"Peak Memory Usage: {mem:.02f} GB")
```
</details>
# Model Performance
## Results (A100 machine)
| Benchmark (Latency) | | |
|----------------------------------|----------------|--------------------------|
| | Qwen/Qwen3-8B | jerryzh168/Qwen3-8B-INT4 |
| latency (batch_size=1) | ?s | ?s (?x speedup) |
<details>
<summary> Reproduce Model Performance Results </summary>
## Setup
Get vllm source code:
```Shell
git clone git@github.com:vllm-project/vllm.git
```
Install vllm
```
VLLM_USE_PRECOMPILED=1 pip install --editable .
```
Run the benchmarks under `vllm` root folder:
## benchmark_latency
### baseline
```Shell
export MODEL=Qwen/Qwen3-8B
python benchmarks/benchmark_latency.py --input-len 256 --output-len 256 --model $MODEL --batch-size 1
```
### INT4
```Shell
export MODEL=jerryzh168/Qwen3-8B-INT4
VLLM_DISABLE_COMPILE_CACHE=1 python benchmarks/benchmark_latency.py --input-len 256 --output-len 256 --model $MODEL --batch-size 1
```
## benchmark_serving
We benchmarked the throughput in a serving environment.
Download sharegpt dataset:
```Shell
wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
```
Other datasets can be found in: https://github.com/vllm-project/vllm/tree/main/benchmarks
Note: you can change the number of prompts to be benchmarked with `--num-prompts` argument for `benchmark_serving` script.
### baseline
Server:
```Shell
export MODEL=Qwen/Qwen3-8B
vllm serve $MODEL --tokenizer $MODEL -O3
```
Client:
```Shell
export MODEL=Qwen/Qwen3-8B
python benchmarks/benchmark_serving.py --backend vllm --dataset-name sharegpt --tokenizer $MODEL --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json --model $MODEL --num-prompts 1
```
### INT4
Server:
```Shell
export MODEL=jerryzh168/Qwen3-8B-INT4
VLLM_DISABLE_COMPILE_CACHE=1 vllm serve $MODEL --tokenizer $MODEL -O3 --pt-load-map-location cuda:0
```
Client:
```Shell
export MODEL=jerryzh168/Qwen3-8B-INT4
python benchmarks/benchmark_serving.py --backend vllm --dataset-name sharegpt --tokenizer $MODEL --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json --model $MODEL --num-prompts 1
```
</details>
# Paper: TorchAO: PyTorch-Native Training-to-Serving Model Optimization
The model's quantization is powered by **TorchAO**, a framework presented in the paper [TorchAO: PyTorch-Native Training-to-Serving Model Optimization](https://huggingface.co/papers/2507.16099).
**Abstract:** We present TorchAO, a PyTorch-native model optimization framework leveraging quantization and sparsity to provide an end-to-end, training-to-serving workflow for AI models. TorchAO supports a variety of popular model optimization techniques, including FP8 quantized training, quantization-aware training (QAT), post-training quantization (PTQ), and 2:4 sparsity, and leverages a novel tensor subclass abstraction to represent a variety of widely-used, backend agnostic low precision data types, including INT4, INT8, FP8, MXFP4, MXFP6, and MXFP8. TorchAO integrates closely with the broader ecosystem at each step of the model optimization pipeline, from pre-training (TorchTitan) to fine-tuning (TorchTune, Axolotl) to serving (HuggingFace, vLLM, SGLang, ExecuTorch), connecting an otherwise fragmented space in a single, unified workflow. TorchAO has enabled recent launches of the quantized Llama 3.2 1B/3B and LlamaGuard3-8B models and is open-source at this https URL .
# Resources
* **Official TorchAO GitHub Repository:** [https://github.com/pytorch/ao](https://github.com/pytorch/ao)
* **TorchAO Documentation:** [https://docs.pytorch.org/ao/stable/index.html](https://docs.pytorch.org/ao/stable/index.html)
# Disclaimer
PyTorch has not performed safety evaluations or red teamed the quantized models. Performance characteristics, outputs, and behaviors may differ from the original models. Users are solely responsible for selecting appropriate use cases, evaluating and mitigating for accuracy, safety, and fairness, ensuring security, and complying with all applicable laws and regulations.
Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the licenses the models are released under, including any limitations of liability or disclaimers of warranties provided therein.
|
duppbuy/blockassist-bc-zealous_galloping_rat_1756498933
|
duppbuy
| 2025-08-29T20:22:42Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"zealous galloping rat",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:22:14Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- zealous galloping rat
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
koloni/blockassist-bc-deadly_graceful_stingray_1756497417
|
koloni
| 2025-08-29T20:22:41Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"deadly graceful stingray",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:22:37Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- deadly graceful stingray
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
liukevin666/blockassist-bc-yawning_striped_cassowary_1756498510
|
liukevin666
| 2025-08-29T20:19:18Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"yawning striped cassowary",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:16:07Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- yawning striped cassowary
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
Sonic-man/blockassist-bc-poisonous_graceful_cow_1756496054
|
Sonic-man
| 2025-08-29T20:16:30Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"poisonous graceful cow",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:16:26Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- poisonous graceful cow
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
bah63843/blockassist-bc-plump_fast_antelope_1756498460
|
bah63843
| 2025-08-29T20:15:10Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"plump fast antelope",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:15:00Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- plump fast antelope
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
nahcrof/stok-sub-1
|
nahcrof
| 2025-08-29T20:08:36Z | 0 | 0 | null |
[
"text-generation",
"en",
"license:mit",
"region:us"
] |
text-generation
| 2025-08-29T04:24:56Z |
---
license: mit
language:
- en
pipeline_tag: text-generation
---
### Model versions
| **Model** |**Parameters**|**RAM used (inference)**|
| :------------------------ | :----------: | :--------------------: |
| stok-0.1 | 3,798 | 6MB |
| stok-0.2 | 4m | 542MB |
| stok-0.3 | 962k | 136MB |
| stok-0.3-large | 28.86m | 4GB |
## Description
stok is a family of models designed to run better at smaller parameter counts and maintain speed despite model size.
stok-sub-1 will contain all versions of the stok model, prior to releasing stok-1.
The goal of creating the stok models is to have models that regardless of size, can be ran incredibly fast on CPUs (including incredibly old ones).
Currently, stok can only contextualize single prompts and will not understand them beyond a single word. So far, each new version (as in 0.1, 0.2, and 0.3)
has brought a new capability to the model. 0.2 gave the model the ability to end it's thought, 0.3 allowed the model to (usually) keep the token prediction within
the context of the prompt. While the model definitely needs a little more help, it's only in version 0.3, there's a lot of work to go (like a new,
less ram intensive, inference engine).
## How to run
First, when using python (more inference engines coming soon) you will need to install the ```run_stok.py``` file. The code for using this will look something like this:
```python
from run_stok import load_model, run_model
# you can replace stok-0.3.json with whichever stok model you want
load_model("stok-0.3.json")
response = run_model("Hello!", max_tokens=100, repetition_penalty=2)
for chunk in response:
print(chunk, end="")
```
this showcases how to use all currently functioning parameters, although max_tokens and repetition penalty are both technically optional.<br><br>
If you'd rather use stokfile (a tool for just testing out the model) here's how you can.
```
python3 stokfile.py -m stok-0.3.json
```
If you want to see the speed of the output, just add -speed to the end, like so...
```
python3 stokfile.py -m stok-0.3.json -speed
```
## Benchmark (SLMB)
| **Model** | **Score** | **Med. Speed** |
| :-----------------------: | :----------: | :------------: |
| stok-0.1 | 1/16 | 318,188 t/s |
| stok-0.2 | 4/16 | 6,101 t/s |
| stok-0.3 | 5/16 | 73,221 t/s |
| stok-0.3-large | 8/16 | 11,184 t/s |
| TinyLLama-v0 (F32) | 0/16 | 1,695 t/s |
| Gemma-3-270m-it (F16) | 12/16 | 46 t/s |
| H2o danube3 500m chat(F32)| 8/16 | 21 t/s |
| Llama 3.2 1B instruct(F16)| 14/16 | 14 t/s |
The CPU used for each test was the AMD Ryzen 7 2700X<br>
RAM: 64GB DDR4<br>
### The SLMB (Small Language Model Benchmark) v1
#### Quick description
This is a very very simple model test, created to test the capabilies of much smaller LLMs. (The answers are included, though they aren't actually needed)
#### The Benchmark
Category 1: elementary math - x/4<br>
what is 2+2 (4)<br>
what is 12+5 (17)<br>
what is 4/2 (2)<br>
what is 3*3 (9)<br>
<br>
<br>
Category 2: math with large numbers - x/4<br>
what is 500+200 (700)<br>
what is 10000+1000 (11000)<br>
what is 100\*100 (10000)<br>
what is 12\*5000 (60000)<br>
<br><br>
Category 3: input variation - x/5<br>
what is 1+1 (2)<br>
what is 1 + 1 (2)<br>
what is 1+ 1 (2)<br>
what is a dog (any answer that matches at least a very basic description of a dog)<br>
What is a dog? (any answer that matches at least a very basic description of a dog)<br>
<br>
Category 4: basic logic - x/2 (2 points for correct, 0 for wrong)<br>
I have three friends (Jeremy, Tyler, and Gabe) Friend #1 is Jeremy, Friend #3 is Tyler, who is friend #2? <br>
(Gabe)<br>
## Conclusion
While stok is definitely (in my opinion) pretty impressive -- especially given it's performance at such small sizes -- it has lots of room to go (also the
benchmark may include more tests in the future)
|
MercuryNex/sdLCM62
|
MercuryNex
| 2025-08-29T20:04:42Z | 0 | 0 |
diffusers
|
[
"diffusers",
"safetensors",
"text-to-image",
"en",
"license:other",
"autotrain_compatible",
"endpoints_compatible",
"diffusers:StableDiffusionPipeline",
"region:us"
] |
text-to-image
| 2025-08-29T20:04:29Z |
---
license: other
language:
- en
library_name: diffusers
pipeline_tag: text-to-image
tags:
- text-to-image
---
Converted from [https://civitai.com/api/download/models/1347172?type=Model&format=SafeTensor&size=pruned&fp=fp16](https://civitai.com/api/download/models/1347172?type=Model&format=SafeTensor&size=pruned&fp=fp16).
|
elmenbillion/blockassist-bc-beaked_sharp_otter_1756496221
|
elmenbillion
| 2025-08-29T20:04:08Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"beaked sharp otter",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:04:03Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- beaked sharp otter
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
AnerYubo/blockassist-bc-snappy_tenacious_eagle_1756497773
|
AnerYubo
| 2025-08-29T20:02:56Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"snappy tenacious eagle",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:02:54Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- snappy tenacious eagle
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
tashfinsami/model_bn
|
tashfinsami
| 2025-08-29T20:02:44Z | 1 | 0 |
diffusers
|
[
"diffusers",
"text-to-image",
"lora",
"diffusers-training",
"stable-diffusion",
"stable-diffusion-diffusers",
"base_model:Kardbord/stable-diffusion-v1-5-unsafe",
"base_model:adapter:Kardbord/stable-diffusion-v1-5-unsafe",
"license:creativeml-openrail-m",
"region:us"
] |
text-to-image
| 2025-08-28T07:20:36Z |
---
base_model: Kardbord/stable-diffusion-v1-5-unsafe
library_name: diffusers
license: creativeml-openrail-m
inference: true
instance_prompt: a derm photo of sks blue naevus lesion
tags:
- text-to-image
- diffusers
- lora
- diffusers-training
- stable-diffusion
- stable-diffusion-diffusers
- text-to-image
- diffusers
- lora
- diffusers-training
- stable-diffusion
- stable-diffusion-diffusers
---
<!-- This model card has been generated automatically according to the information the training script had access to. You
should probably proofread and complete it, then remove this comment. -->
# LoRA DreamBooth - tashfinsami/model_bn
These are LoRA adaption weights for Kardbord/stable-diffusion-v1-5-unsafe. The weights were trained on a derm photo of sks blue naevus lesion using [DreamBooth](https://dreambooth.github.io/). You can find some example images in the following.
LoRA for the text encoder was enabled: True.
## Intended uses & limitations
#### How to use
```python
# TODO: add an example code snippet for running this diffusion pipeline
```
#### Limitations and bias
[TODO: provide examples of latent issues and potential remediations]
## Training details
[TODO: describe the data used to train the model]
|
boonpertou/blockassist-bc-small_durable_jaguar_1756497689
|
boonpertou
| 2025-08-29T20:02:19Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"small durable jaguar",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T20:01:29Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- small durable jaguar
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
weathermanj/Nemotron-nano-9b-fp8
|
weathermanj
| 2025-08-29T20:01:36Z | 753 | 6 |
transformers
|
[
"transformers",
"safetensors",
"nvidia",
"nemotron",
"fp8",
"quantized",
"quantization",
"efficient-inference",
"hybrid-mamba-transformer",
"text-generation",
"conversational",
"en",
"base_model:nvidia/NVIDIA-Nemotron-Nano-9B-v2",
"base_model:finetune:nvidia/NVIDIA-Nemotron-Nano-9B-v2",
"license:other",
"endpoints_compatible",
"region:us"
] |
text-generation
| 2025-08-21T03:23:17Z |
---
language:
- en
license: other
license_name: nvidia-open-model-license
license_link: >-
https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf
library_name: transformers
tags:
- nvidia
- nemotron
- fp8
- quantized
- quantization
- efficient-inference
- hybrid-mamba-transformer
base_model: nvidia/NVIDIA-Nemotron-Nano-9B-v2
model_type: nemotron_h
pipeline_tag: text-generation
inference: true
---
# NVIDIA-Nemotron-Nano-9B-v2-FP8
**FP8 Quantized by jwjohns | Emendat.io**
This is an FP8-quantized version of [nvidia/NVIDIA-Nemotron-Nano-9B-v2](https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2) optimized for efficient inference with significant memory reduction while preserving model quality.
## Model Overview
- **Base Model**: [nvidia/NVIDIA-Nemotron-Nano-9B-v2](https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2)
- **Model Architecture**: Hybrid Mamba-Transformer
- **Parameters**: 8.89B (effective)
- **Model Size**: 9.48 GB (vs 17.78 GB original)
- **Compression**: 1.88x smaller (46.7% size reduction)
- **Quantization**: FP8 E4M3 format with smart precision preservation
- **Context Length**: Up to 128K tokens
- **License**: [NVIDIA Open Model License](https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf)
## Quantization Details
This model was quantized using a custom FP8 conversion process that:
- Converts linear layer weights to FP8 E4M3 format (1 byte per parameter)
- Preserves embeddings, layer norms, and biases in BF16 for stability
- Maintains the hybrid Mamba-Transformer architecture integrity
- Creates actual quantized weights (not runtime quantization)
### Technical Specifications
- **Original Size**: 17.78 GB
- **Quantized Size**: 9.48 GB
- **Compression Ratio**: 1.88x
- **Memory Reduction**: 46.7%
- **Conversion Method**: Direct safetensors FP8 weight conversion
- **Preserved Layers**: Embeddings, LayerNorms, Biases (BF16)
- **Quantized Layers**: Linear/MLP weights (FP8 E4M3)
## Architecture Details
The hybrid architecture is fully preserved:
- **27 Mamba2 layers**: Efficient O(n) sequence processing
- **4 Attention layers**: Complex reasoning and context understanding
- **25 MLP layers**: Feed-forward processing
- **Total**: 56 layers optimized for both efficiency and capability
## Usage
### Recommended: vLLM (Optimal Performance)
```python
from vllm import LLM, SamplingParams
# Load the FP8 quantized model
model = LLM(
model="weathermanj/nemotron-nano-9b-fp8",
trust_remote_code=True,
dtype="auto" # Will auto-detect FP8 format
)
# Generate with streaming support
prompts = ["Explain the benefits of hybrid Mamba-Transformer architectures."]
sampling_params = SamplingParams(
temperature=0.7,
top_p=0.9,
max_tokens=256
)
outputs = model.generate(prompts, sampling_params)
for output in outputs:
print(output.outputs[0].text)
```
### Alternative: Transformers
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# Load the quantized model
model = AutoModelForCausalLM.from_pretrained(
"weathermanj/nemotron-nano-9b-fp8",
trust_remote_code=True,
torch_dtype="auto",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("weathermanj/nemotron-nano-9b-fp8")
# Generate text
inputs = tokenizer("How does FP8 quantization improve AI efficiency?", return_tensors="pt")
with torch.no_grad():
outputs = model.generate(**inputs, max_length=200, do_sample=True, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
```
## Performance Comparison
| Metric | Original BF16 | FP8 Quantized | Improvement |
|--------|---------------|---------------|-------------|
| **Model Size** | 17.78 GB | 9.48 GB | 46.7% smaller |
| **Memory Usage** | ~28 GB | ~19 GB | 32% reduction |
| **VRAM Required** | 20+ GB | 12+ GB | More accessible |
| **Quality Loss** | 0% | <2% | Minimal degradation |
| **Inference Speed** | Baseline | Up to 1.5x | Faster on supported HW |
## Hardware Requirements
### Optimal Performance
- **H100, A100**: Native FP8 support for maximum efficiency
- **Ada Lovelace (RTX 4090)**: Excellent FP8 performance
- **Memory**: 12GB+ VRAM for inference
### Compatible Hardware
- **RTX 3080/3090**: Software FP8 emulation
- **V100, T4**: Falls back to BF16 with memory benefits
- **Memory**: 14GB+ VRAM recommended
## Supported Languages
The model maintains full multilingual capabilities:
- English
- German
- Spanish
- French
- Italian
- Japanese
## Use Cases
This quantized model is ideal for:
- **Production deployments** requiring memory efficiency
- **Edge inference** on resource-constrained hardware
- **High-throughput serving** with vLLM
- **Development/research** with reduced VRAM requirements
- **Hybrid reasoning tasks** leveraging Mamba+Attention architecture
## Chat Template
The model uses the standard chat template format:
```
<extra_id_0>System
{system_message}
<extra_id_1>User
{user_message}
<extra_id_1>Assistant
{assistant_message}
```
## Limitations
- Slight quality degradation possible in edge cases (<2%)
- Requires FP8-compatible hardware for optimal performance
- May fall back to higher precision on older GPUs
## Technical Notes
- **Quantization preserves** the model's reasoning capabilities and multilingual performance
- **Hybrid architecture benefits** are maintained (fast Mamba layers + powerful attention)
- **Compatible** with existing NVIDIA Nemotron inference pipelines
- **Safetensors format** ensures safe and efficient loading
## Citation
```bibtex
@software{nemotron_fp8_quantized,
title={NVIDIA-Nemotron-Nano-9B-v2-FP8: Efficient FP8 Quantization},
author={jwjohns},
organization={Emendat.io},
year={2025},
url={https://huggingface.co/weathermanj/nvidia-nemotron-nano-9b-v2-fp8},
note={FP8 quantized version of NVIDIA Nemotron-Nano-9B-v2}
}
@article{nvidia2024nemotron,
title={Nemotron-4 Technical Report},
author={NVIDIA},
year={2024},
url={https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2}
}
```
## License
This quantized model inherits the [NVIDIA Open Model License](https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf) from the original model.
---
## Model Tracking & Attribution
### Quantization Details
- **Quantized by**: jwjohns (Emendat.io)
- **Quantization Date**: 2025-08-21
- **Original Model**: [nvidia/NVIDIA-Nemotron-Nano-9B-v2](https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2)
- **Quantization Method**: Custom FP8 weight conversion
- **Framework**: Direct safetensors manipulation with PyTorch FP8 support
- **Repository**: [Nemotron-Ozempic Project](https://github.com/jwjohns/nemotron-ozempic)
### Conversion Pipeline
1. **Source**: NVIDIA Nemotron-Nano-9B-v2 (locally cached)
2. **Conversion**: Custom FP8 E4M3 weight conversion script
3. **Preservation**: Smart layer selection (embeddings/norms → BF16, weights → FP8)
4. **Validation**: Safetensors format integrity check
5. **Upload**: HuggingFace Hub with full metadata
### Model Lineage
```
nvidia/NVIDIA-Nemotron-Nano-9B-v2 (Base Model)
↓ (FP8 Quantization by jwjohns)
weathermanj/nvidia-nemotron-nano-9b-v2-fp8 (This Model)
```
### Usage Tracking
If you use this model, please cite both the original NVIDIA work and the quantization:
```bibtex
@software{nvidia_nemotron_fp8,
title={NVIDIA-Nemotron-Nano-9B-v2-FP8},
author={jwjohns},
organization={Emendat.io},
year={2025},
url={https://huggingface.co/weathermanj/Nemotron-nano-9b-fp8},
note={FP8 quantized version of NVIDIA Nemotron-Nano-9B-v2},
baseModel={nvidia/NVIDIA-Nemotron-Nano-9B-v2}
}
@article{nvidia2024nemotron,
title={Nemotron-4 Technical Report},
author={NVIDIA},
year={2024},
url={https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2}
}
```
### Quality Assurance
- ✅ **Weights verified**: All FP8 conversions validated
- ✅ **Format integrity**: Safetensors format preserved
- ✅ **Architecture preserved**: Hybrid Mamba-Transformer intact
- ✅ **Tokenizer compatibility**: Original tokenizer maintained
- ✅ **Config validation**: Quantization metadata added
- ✅ **License compliance**: NVIDIA Open Model License respected
## Model Tracking & Attribution
### Quantization Details
- **Quantized by**: jwjohns (Emendat.io)
- **Quantization Date**: 2025-08-21
- **Original Model**: [nvidia/NVIDIA-Nemotron-Nano-9B-v2](https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2)
- **Quantization Method**: Custom FP8 weight conversion
- **Framework**: Direct safetensors manipulation with PyTorch FP8 support
- **Repository**: [Nemotron-Ozempic Project](https://github.com/jwjohns/nemotron-ozempic)
### Conversion Pipeline
1. **Source**: NVIDIA Nemotron-Nano-9B-v2 (locally cached)
2. **Conversion**: Custom FP8 E4M3 weight conversion script
3. **Preservation**: Smart layer selection (embeddings/norms → BF16, weights → FP8)
4. **Validation**: Safetensors format integrity check
5. **Upload**: HuggingFace Hub with full metadata
### Model Lineage
```
nvidia/NVIDIA-Nemotron-Nano-9B-v2 (Base Model)
↓ (FP8 Quantization by jwjohns)
weathermanj/nvidia-nemotron-nano-9b-v2-fp8 (This Model)
```
### Usage Tracking
If you use this model, please cite both the original NVIDIA work and the quantization:
```bibtex
@software{nvidia_nemotron_fp8,
title={NVIDIA-Nemotron-Nano-9B-v2-FP8},
author={jwjohns},
organization={Emendat.io},
year={2025},
url={https://huggingface.co/weathermanj/Nemotron-nano-9b-fp8},
note={FP8 quantized version of NVIDIA Nemotron-Nano-9B-v2},
baseModel={nvidia/NVIDIA-Nemotron-Nano-9B-v2}
}
@article{nvidia2024nemotron,
title={Nemotron-4 Technical Report},
author={NVIDIA},
year={2024},
url={https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2}
}
```
### Quality Assurance
- ✅ **Weights verified**: All FP8 conversions validated
- ✅ **Format integrity**: Safetensors format preserved
- ✅ **Architecture preserved**: Hybrid Mamba-Transformer intact
- ✅ **Tokenizer compatibility**: Original tokenizer maintained
- ✅ **Config validation**: Quantization metadata added
- ✅ **License compliance**: NVIDIA Open Model License respected
**Quantization by jwjohns | Emendat.io • Base model by NVIDIA**
This FP8 quantization demonstrates successful compression of hybrid Mamba-Transformer architectures while maintaining the benefits of both efficient sequence processing and powerful reasoning capabilities.
|
brknnode/blockassist-bc-wise_invisible_cat_1756497521
|
brknnode
| 2025-08-29T20:00:32Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"wise invisible cat",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:59:31Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- wise invisible cat
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
duppbuy/blockassist-bc-flapping_unseen_scorpion_1756497585
|
duppbuy
| 2025-08-29T20:00:29Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"flapping unseen scorpion",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:59:46Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- flapping unseen scorpion
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
mooperyou/blockassist-bc-beaked_frisky_ox_1756497526
|
mooperyou
| 2025-08-29T19:59:25Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"beaked frisky ox",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:58:47Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- beaked frisky ox
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
mooperyou/blockassist-bc-beaked_frisky_ox_1756497405
|
mooperyou
| 2025-08-29T19:57:27Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"beaked frisky ox",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:56:46Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- beaked frisky ox
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
seanchen11235/blockassist-bc-monstrous_ravenous_quail_1756493549
|
seanchen11235
| 2025-08-29T19:56:30Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"monstrous ravenous quail",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:55:56Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- monstrous ravenous quail
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
mooperyou/blockassist-bc-beaked_frisky_ox_1756497288
|
mooperyou
| 2025-08-29T19:55:24Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"beaked frisky ox",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:54:48Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- beaked frisky ox
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
Bisher/laced-tree-2-merged-f16
|
Bisher
| 2025-08-29T19:55:15Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"gemma3",
"image-text-to-text",
"text-generation-inference",
"unsloth",
"conversational",
"en",
"base_model:unsloth/gemma-3-4b-it",
"base_model:finetune:unsloth/gemma-3-4b-it",
"license:apache-2.0",
"endpoints_compatible",
"region:us"
] |
image-text-to-text
| 2025-08-29T19:54:22Z |
---
base_model: unsloth/gemma-3-4b-it
tags:
- text-generation-inference
- transformers
- unsloth
- gemma3
license: apache-2.0
language:
- en
---
# Uploaded finetuned model
- **Developed by:** Bisher
- **License:** apache-2.0
- **Finetuned from model :** unsloth/gemma-3-4b-it
This gemma3 model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
|
helmutsukocok/blockassist-bc-loud_scavenging_kangaroo_1756495695
|
helmutsukocok
| 2025-08-29T19:53:22Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"loud scavenging kangaroo",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:53:17Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- loud scavenging kangaroo
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
Stasonelison/blockassist-bc-howling_powerful_aardvark_1756497130
|
Stasonelison
| 2025-08-29T19:53:11Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"howling powerful aardvark",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:52:49Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- howling powerful aardvark
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
Bisher/laced-tree-2
|
Bisher
| 2025-08-29T19:52:43Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"generated_from_trainer",
"unsloth",
"sft",
"trl",
"base_model:unsloth/gemma-3-4b-it",
"base_model:finetune:unsloth/gemma-3-4b-it",
"endpoints_compatible",
"region:us"
] | null | 2025-08-29T18:14:11Z |
---
base_model: unsloth/gemma-3-4b-it
library_name: transformers
model_name: laced-tree-2
tags:
- generated_from_trainer
- unsloth
- sft
- trl
licence: license
---
# Model Card for laced-tree-2
This model is a fine-tuned version of [unsloth/gemma-3-4b-it](https://huggingface.co/unsloth/gemma-3-4b-it).
It has been trained using [TRL](https://github.com/huggingface/trl).
## Quick start
```python
from transformers import pipeline
question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="Bisher/laced-tree-2", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])
```
## Training procedure
[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>](https://wandb.ai/bishertello-/Gemma-H200/runs/tlq9xwn1)
This model was trained with SFT.
### Framework versions
- TRL: 0.21.0
- Transformers: 4.55.4
- Pytorch: 2.8.0
- Datasets: 3.6.0
- Tokenizers: 0.21.4
## Citations
Cite TRL as:
```bibtex
@misc{vonwerra2022trl,
title = {{TRL: Transformer Reinforcement Learning}},
author = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
year = 2020,
journal = {GitHub repository},
publisher = {GitHub},
howpublished = {\url{https://github.com/huggingface/trl}}
}
```
|
dgambettaphd/M_llm2_run1_gen8_X_doc1000_synt64_lr1e-04_acm_SYNLAST
|
dgambettaphd
| 2025-08-29T19:52:14Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"unsloth",
"arxiv:1910.09700",
"endpoints_compatible",
"region:us"
] | null | 2025-08-29T19:51:51Z |
---
library_name: transformers
tags:
- unsloth
---
# Model Card for Model ID
<!-- Provide a quick summary of what the model is/does. -->
## Model Details
### Model Description
<!-- Provide a longer summary of what this model is. -->
This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
- **Developed by:** [More Information Needed]
- **Funded by [optional]:** [More Information Needed]
- **Shared by [optional]:** [More Information Needed]
- **Model type:** [More Information Needed]
- **Language(s) (NLP):** [More Information Needed]
- **License:** [More Information Needed]
- **Finetuned from model [optional]:** [More Information Needed]
### Model Sources [optional]
<!-- Provide the basic links for the model. -->
- **Repository:** [More Information Needed]
- **Paper [optional]:** [More Information Needed]
- **Demo [optional]:** [More Information Needed]
## Uses
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
### Direct Use
<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
[More Information Needed]
### Downstream Use [optional]
<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
[More Information Needed]
### Out-of-Scope Use
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
[More Information Needed]
## Bias, Risks, and Limitations
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
[More Information Needed]
### Recommendations
<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
## How to Get Started with the Model
Use the code below to get started with the model.
[More Information Needed]
## Training Details
### Training Data
<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
[More Information Needed]
### Training Procedure
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
#### Preprocessing [optional]
[More Information Needed]
#### Training Hyperparameters
- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
#### Speeds, Sizes, Times [optional]
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
[More Information Needed]
## Evaluation
<!-- This section describes the evaluation protocols and provides the results. -->
### Testing Data, Factors & Metrics
#### Testing Data
<!-- This should link to a Dataset Card if possible. -->
[More Information Needed]
#### Factors
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
[More Information Needed]
#### Metrics
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
[More Information Needed]
### Results
[More Information Needed]
#### Summary
## Model Examination [optional]
<!-- Relevant interpretability work for the model goes here -->
[More Information Needed]
## Environmental Impact
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
- **Hardware Type:** [More Information Needed]
- **Hours used:** [More Information Needed]
- **Cloud Provider:** [More Information Needed]
- **Compute Region:** [More Information Needed]
- **Carbon Emitted:** [More Information Needed]
## Technical Specifications [optional]
### Model Architecture and Objective
[More Information Needed]
### Compute Infrastructure
[More Information Needed]
#### Hardware
[More Information Needed]
#### Software
[More Information Needed]
## Citation [optional]
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
**BibTeX:**
[More Information Needed]
**APA:**
[More Information Needed]
## Glossary [optional]
<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
[More Information Needed]
## More Information [optional]
[More Information Needed]
## Model Card Authors [optional]
[More Information Needed]
## Model Card Contact
[More Information Needed]
|
poeryouy/blockassist-bc-skilled_omnivorous_elephant_1756497074
|
poeryouy
| 2025-08-29T19:51:56Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"skilled omnivorous elephant",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:51:15Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- skilled omnivorous elephant
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
huggingtoots/TheDrummer-GLM-Steam-106B-A12B-v1-MLX-8Bit
|
huggingtoots
| 2025-08-29T19:50:34Z | 0 | 0 |
mlx
|
[
"mlx",
"safetensors",
"glm4_moe",
"base_model:TheDrummer/GLM-Steam-106B-A12B-v1",
"base_model:quantized:TheDrummer/GLM-Steam-106B-A12B-v1",
"8-bit",
"region:us"
] | null | 2025-08-29T19:27:56Z |
---
base_model: TheDrummer/GLM-Steam-106B-A12B-v1
tags:
- mlx
---
# huggingtoots/TheDrummer-GLM-Steam-106B-A12B-v1-MLX-8Bit
The Model [huggingtoots/TheDrummer-GLM-Steam-106B-A12B-v1-MLX-8Bit](https://huggingface.co/huggingtoots/TheDrummer-GLM-Steam-106B-A12B-v1-MLX-8Bit) was converted to MLX format from [TheDrummer/GLM-Steam-106B-A12B-v1](https://huggingface.co/TheDrummer/GLM-Steam-106B-A12B-v1) using mlx-lm version **0.26.4**.
🦛 <span style="color:#800080">If you want a free consulting session, </span>[fill out this form](https://forms.gle/xM9gw1urhypC4bWS6) <span style="color:#800080">to get in touch!</span> 🤗
## Use with mlx
```bash
pip install mlx-lm
```
```python
from mlx_lm import load, generate
model, tokenizer = load("huggingtoots/GLM-Steam-106B-A12B-v1-mlx-8Bit")
prompt="hello"
if hasattr(tokenizer, "apply_chat_template") and tokenizer.chat_template is not None:
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
response = generate(model, tokenizer, prompt=prompt, verbose=True)
```
|
Amama02/pinterest-personality-keywords-reduced-epochs
|
Amama02
| 2025-08-29T19:50:21Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"t5",
"text2text-generation",
"pinterest",
"keywords",
"personality",
"fine-tuned",
"lora",
"flan-t5",
"en",
"base_model:google/flan-t5-base",
"base_model:adapter:google/flan-t5-base",
"license:apache-2.0",
"text-generation-inference",
"endpoints_compatible",
"region:us"
] |
text-generation
| 2025-08-29T19:49:58Z |
---
language: en
license: apache-2.0
base_model: google/flan-t5-base
tags:
- text2text-generation
- pinterest
- keywords
- personality
- fine-tuned
- lora
- flan-t5
library_name: transformers
pipeline_tag: text2text-generation
widget:
- text: "Generate Pinterest keywords for Cleopatra - Culture: Egyptian | Role: Royalty | Period: Ancient Egypt - Keywords should be visual, searchable on Pinterest, and capture their aesthetic essence. The Culture, Role, Period and bio give important information about the personality. Take them into account when generating keywords"
example_title: "Cleopatra Keywords"
- text: "Generate Pinterest keywords for Leonardo da Vinci - Culture: Italian | Role: Polymath | Period: Renaissance - Keywords should be visual, searchable on Pinterest, and capture their aesthetic essence. The Culture, Role, Period and bio give important information about the personality. Take them into account when generating keywords"
example_title: "Leonardo da Vinci Keywords"
---
# Pinterest Personality Keywords Generator
🎨 **Fine-tuned FLAN-T5 model for generating Pinterest-optimized keywords for historical and fictional personalities.**
This model was fine-tuned using LoRA (Low-Rank Adaptation) to generate visually appealing, searchable Pinterest keywords based on personality information.
## 🚀 Quick Start
### Using Transformers Pipeline
```python
from transformers import pipeline
# Load the model
generator = pipeline("text2text-generation", model="Amama02/pinterest-personality-keywords-reduced-epochs")
# Generate keywords
input_text = "Generate Pinterest keywords for Marie Curie - Culture: Polish-French | Role: Scientist | Period: Early 20th Century - Keywords should be visual, searchable on Pinterest, and capture their aesthetic essence. The Culture, Role, Period and bio give important information about the personality. Take them into account when generating keywords"
result = generator(
input_text,
max_length=300,
num_beams=8,
temperature=0.9,
do_sample=True,
top_p=0.95,
repetition_penalty=2.0,
length_penalty=1.2,
no_repeat_ngram_size=2
)
print(result[0]['generated_text'])
```
### Using Direct Model Loading
```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("Amama02/pinterest-personality-keywords-reduced-epochs")
model = AutoModelForSeq2SeqLM.from_pretrained("Amama02/pinterest-personality-keywords-reduced-epochs")
# Prepare input
input_text = "Generate Pinterest keywords for Frida Kahlo - Culture: Mexican | Role: Artist | Period: 20th Century - Keywords should be visual, searchable on Pinterest, and capture their aesthetic essence. The Culture, Role, Period and bio give important information about the personality. Take them into account when generating keywords"
# Tokenize and generate
inputs = tokenizer(input_text, return_tensors="pt", max_length=256, truncation=True)
outputs = model.generate(
**inputs,
max_length=300,
num_beams=8,
temperature=0.9,
do_sample=True,
top_p=0.95,
repetition_penalty=2.0,
length_penalty=1.2,
early_stopping=True,
no_repeat_ngram_size=2
)
keywords = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(keywords)
```
## 📝 Input Format
The model expects input in this specific format:
```
Generate Pinterest keywords for [PERSONALITY_NAME] - Culture: [CULTURE] | Role: [ROLE] | Period: [TIME_PERIOD] | Bio: [BIOGRAPHY] - Keywords should be visual, searchable on Pinterest, and capture their aesthetic essence. The Culture, Role, Period and bio give important information about the personality. Take them into account when generating keywords
```
### Required Fields:
- **PERSONALITY_NAME**: Name of the person
- **Culture**: Cultural background or nationality
- **Role**: Profession, title, or main role
- **Period**: Historical time period
- **Bio**: (Optional) Brief biography
## 🎯 Example Outputs
| Input | Generated Keywords |
|-------|-------------------|
| **Cleopatra** (Egyptian Royalty, Ancient Egypt) | "Egyptian queen aesthetic, ancient Egypt fashion, Cleopatra makeup, pharaoh style, golden jewelry, Egyptian mythology, ancient beauty, royal Egyptian, hieroglyphics, Egyptian art" |
| **Leonardo da Vinci** (Italian Polymath, Renaissance) | "Renaissance art, Italian genius, classical paintings, Renaissance fashion, vintage sketches, Italian Renaissance, Renaissance architecture, classical art history" |
| **Marie Curie** (Polish-French Scientist, Early 20th Century) | "vintage science, female scientist aesthetic, laboratory vintage, early 1900s fashion, women in science, vintage academic, scientific discovery, vintage portraits" |
## ⚙️ Generation Parameters
The model is optimized with these generation settings:
- **max_length**: 300
- **num_beams**: 8
- **temperature**: 0.9
- **top_p**: 0.95
- **repetition_penalty**: 2.0
- **length_penalty**: 1.2
- **no_repeat_ngram_size**: 2
## 🔧 Technical Details
- **Base Model**: google/flan-t5-base
- **Fine-tuning Method**: LoRA (Low-Rank Adaptation)
- **LoRA Rank**: 16
- **Target Modules**: ["q", "v", "k", "o", "wi", "wo"]
- **Training Data**: Historical and fictional personalities dataset
- **Task**: Seq2Seq text generation
The model has been optimized for:
- ✅ **Visual Keywords**: Generates terms that work well for image searches
- ✅ **Pinterest Optimization**: Keywords tailored for Pinterest's search algorithm
- ✅ **Cultural Sensitivity**: Respects cultural context and historical accuracy
- ✅ **Diversity**: Produces varied and creative keyword combinations
## 🚫 Limitations
- Specifically designed for Pinterest keyword generation
- May not perform well on other text generation tasks
- Limited to personalities with sufficient historical/cultural context
- Requires specific input format for optimal results
|
motza0025/blockassist-bc-sturdy_leaping_jaguar_1756495502
|
motza0025
| 2025-08-29T19:48:14Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"sturdy leaping jaguar",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:47:53Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- sturdy leaping jaguar
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
acidjp/blockassist-bc-pesty_extinct_prawn_1756494349
|
acidjp
| 2025-08-29T19:45:54Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"pesty extinct prawn",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:45:50Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- pesty extinct prawn
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
bah63843/blockassist-bc-plump_fast_antelope_1756496610
|
bah63843
| 2025-08-29T19:44:28Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"plump fast antelope",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:44:07Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- plump fast antelope
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
AlignmentResearch/Llama-2-7b-chat-hf-gsm8k-lora-reference
|
AlignmentResearch
| 2025-08-29T19:43:31Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"arxiv:1910.09700",
"endpoints_compatible",
"region:us"
] | null | 2025-08-29T19:43:08Z |
---
library_name: transformers
tags: []
---
# Model Card for Model ID
<!-- Provide a quick summary of what the model is/does. -->
## Model Details
### Model Description
<!-- Provide a longer summary of what this model is. -->
This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
- **Developed by:** [More Information Needed]
- **Funded by [optional]:** [More Information Needed]
- **Shared by [optional]:** [More Information Needed]
- **Model type:** [More Information Needed]
- **Language(s) (NLP):** [More Information Needed]
- **License:** [More Information Needed]
- **Finetuned from model [optional]:** [More Information Needed]
### Model Sources [optional]
<!-- Provide the basic links for the model. -->
- **Repository:** [More Information Needed]
- **Paper [optional]:** [More Information Needed]
- **Demo [optional]:** [More Information Needed]
## Uses
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
### Direct Use
<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
[More Information Needed]
### Downstream Use [optional]
<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
[More Information Needed]
### Out-of-Scope Use
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
[More Information Needed]
## Bias, Risks, and Limitations
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
[More Information Needed]
### Recommendations
<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
## How to Get Started with the Model
Use the code below to get started with the model.
[More Information Needed]
## Training Details
### Training Data
<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
[More Information Needed]
### Training Procedure
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
#### Preprocessing [optional]
[More Information Needed]
#### Training Hyperparameters
- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
#### Speeds, Sizes, Times [optional]
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
[More Information Needed]
## Evaluation
<!-- This section describes the evaluation protocols and provides the results. -->
### Testing Data, Factors & Metrics
#### Testing Data
<!-- This should link to a Dataset Card if possible. -->
[More Information Needed]
#### Factors
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
[More Information Needed]
#### Metrics
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
[More Information Needed]
### Results
[More Information Needed]
#### Summary
## Model Examination [optional]
<!-- Relevant interpretability work for the model goes here -->
[More Information Needed]
## Environmental Impact
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
- **Hardware Type:** [More Information Needed]
- **Hours used:** [More Information Needed]
- **Cloud Provider:** [More Information Needed]
- **Compute Region:** [More Information Needed]
- **Carbon Emitted:** [More Information Needed]
## Technical Specifications [optional]
### Model Architecture and Objective
[More Information Needed]
### Compute Infrastructure
[More Information Needed]
#### Hardware
[More Information Needed]
#### Software
[More Information Needed]
## Citation [optional]
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
**BibTeX:**
[More Information Needed]
**APA:**
[More Information Needed]
## Glossary [optional]
<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
[More Information Needed]
## More Information [optional]
[More Information Needed]
## Model Card Authors [optional]
[More Information Needed]
## Model Card Contact
[More Information Needed]
|
mooperyou/blockassist-bc-alert_melodic_swan_1756496533
|
mooperyou
| 2025-08-29T19:43:06Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"alert melodic swan",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:42:13Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- alert melodic swan
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
poeryouy/blockassist-bc-skilled_omnivorous_elephant_1756496446
|
poeryouy
| 2025-08-29T19:41:28Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"skilled omnivorous elephant",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:40:47Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- skilled omnivorous elephant
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
ilintar/NVIDIA-Nemotron-Nano-9B-v2-GGUF
|
ilintar
| 2025-08-29T19:41:07Z | 0 | 1 | null |
[
"gguf",
"base_model:nvidia/NVIDIA-Nemotron-Nano-9B-v2",
"base_model:quantized:nvidia/NVIDIA-Nemotron-Nano-9B-v2",
"license:other",
"region:us"
] | null | 2025-08-29T14:41:24Z |
---
license: other
license_name: nvidia-open-model-license
license_link: >-
https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/
base_model:
- nvidia/NVIDIA-Nemotron-Nano-9B-v2
---
IMatrix GGUFs calibrated on https://huggingface.co/datasets/eaddario/imatrix-calibration/tree/main combined_all_small set.
Note: Due to the nonstandard tensor sizes, some quantization types do not make sense. For example, due to fallbacks IQ2_M is just 300MB smaller than IQ4_NL. Thus, I only upload the quantizations that actually made sense.
|
Stasonelison/blockassist-bc-howling_powerful_aardvark_1756496317
|
Stasonelison
| 2025-08-29T19:39:25Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"howling powerful aardvark",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:39:15Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- howling powerful aardvark
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
brknnode1/blockassist-bc-lethal_feathered_worm_1756496308
|
brknnode1
| 2025-08-29T19:39:23Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"lethal feathered worm",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:39:02Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- lethal feathered worm
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
DaanBooy/review-sentiment-distilbert-base-uncased
|
DaanBooy
| 2025-08-29T19:39:14Z | 0 | 0 | null |
[
"safetensors",
"distilbert",
"license:apache-2.0",
"region:us"
] | null | 2025-08-29T19:27:41Z |
---
license: apache-2.0
---
|
huwhitememes/gavinnewsom_v1-wan2.1
|
huwhitememes
| 2025-08-29T19:38:57Z | 0 | 0 | null |
[
"video",
"lora",
"wan2.1",
"gavin-newsom",
"generative-video",
"huwhitememes",
"Meme King Studio",
"Green Frog Labs",
"text-to-video",
"base_model:Wan-AI/Wan2.1-T2V-14B",
"base_model:adapter:Wan-AI/Wan2.1-T2V-14B",
"license:apache-2.0",
"region:us"
] |
text-to-video
| 2025-08-29T15:32:31Z |
---
license: apache-2.0
base_model: Wan-AI/Wan2.1-T2V-14B
tags:
- video
- lora
- wan2.1
- gavin-newsom
- generative-video
- huwhitememes
- Meme King Studio
- Green Frog Labs
pipeline_tag: text-to-video
---
# Gavin Newsom LoRA for Wan2.1 (T2V-14B)
This is a custom-trained **LoRA (Low-Rank Adapter)** for **Wan2.1 T2V-14B**, fine-tuned on 24 high-resolution and upscaled images of Gavin Newsom. Designed for **generative video models**, it supports cinematic, political, and meme-style outputs. Trained by [@huwhitememes](https://huggingface.co/huwhitememes) using the [WaveSpeedAI LoRA Trainer](https://wavespeed.ai/models/wavespeed-ai/wan-2.1-14b-lora-trainer).
## 🎯 Use Cases
- Political memes and satire
- Cinematic shitpost visuals
- Social media propaganda videos
## 🔧 Training Details
- **Base Model**: Wan-AI/Wan2.1-T2V-14B
- **Steps**: 2000
- **LoRA Rank**: 32
- **Learning Rate**: 0.0001
- **GPU**: Nvidia H100 (WaveSpeedAI)
- **Image Count**: 24 (upscaled, face-centered, curated)
- **Trigger Word**: `G4vin N3wsom` (recommended in prompt)
---
## 🧠 Creator
Created and uploaded by [@huwhitememes](https://x.com/huwhitememes)
Part of the Meme King Studio / Green Frog Labs creative ecosystem.
## ⚠️ Legal & Fair Use
This model was trained on public photos of a well-known public figure. It is intended for creative expression, satire, parody, education, and research. Please use responsibly and in accordance with your platform’s content guidelines and applicable laws.
## 🧪 Example Usage Prompt
```text
G4vin N3wsom walking through smoke and debris in slow motion, wearing a sharp black suit, chaos around him, golden hour lighting, cinematic movie trailer style, ultra-detailed | wan-style
|
mooperyou/blockassist-bc-alert_melodic_swan_1756496201
|
mooperyou
| 2025-08-29T19:37:19Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"alert melodic swan",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:36:42Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- alert melodic swan
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
jopergil/blockassist-bc-territorial_ferocious_chimpanzee_1756496023
|
jopergil
| 2025-08-29T19:34:16Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"territorial ferocious chimpanzee",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:33:44Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- territorial ferocious chimpanzee
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
bah63843/blockassist-bc-plump_fast_antelope_1756495974
|
bah63843
| 2025-08-29T19:33:45Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"plump fast antelope",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T19:33:36Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- plump fast antelope
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
SreyaDvn/savedModelLebel1
|
SreyaDvn
| 2025-08-29T19:32:17Z | 0 | 0 | null |
[
"safetensors",
"bert",
"license:apache-2.0",
"region:us"
] | null | 2025-08-28T22:03:27Z |
---
license: apache-2.0
---
|
liukevin666/blockassist-bc-yawning_striped_cassowary_1756491874
|
liukevin666
| 2025-08-29T18:25:50Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"yawning striped cassowary",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:25:30Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- yawning striped cassowary
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
goptouy/blockassist-bc-carnivorous_tawny_stingray_1756491916
|
goptouy
| 2025-08-29T18:25:33Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"carnivorous tawny stingray",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:25:16Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- carnivorous tawny stingray
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
barflyman/bert-base-german-finetuned-ler-onnx
|
barflyman
| 2025-08-29T18:23:51Z | 0 | 0 | null |
[
"onnx",
"bert",
"license:apache-2.0",
"region:us"
] | null | 2025-08-29T18:18:12Z |
---
license: apache-2.0
---
BERT-base-German-finetuned-LER-ONNX
This is a converted ONNX version of the mrm8488/bert-base-german-finetuned-ler model, fine-tuned for Legal Entity Recognition (LER) in German texts.
License
The original model is licensed under Apache-2.0. This converted version inherits the same license. You can use and redistribute this model in accordance with the terms of the Apache-2.0 license.
|
Drgaa/Qwen2.5-0.5B-Instruct-Gensyn-Swarm-mammalian_barky_rooster
|
Drgaa
| 2025-08-29T18:23:18Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"qwen2",
"text-generation",
"rl-swarm",
"genrl-swarm",
"grpo",
"gensyn",
"I am mammalian_barky_rooster",
"arxiv:1910.09700",
"autotrain_compatible",
"text-generation-inference",
"endpoints_compatible",
"region:us"
] |
text-generation
| 2025-08-29T12:57:13Z |
---
library_name: transformers
tags:
- rl-swarm
- genrl-swarm
- grpo
- gensyn
- I am mammalian_barky_rooster
---
# Model Card for Model ID
<!-- Provide a quick summary of what the model is/does. -->
## Model Details
### Model Description
<!-- Provide a longer summary of what this model is. -->
This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
- **Developed by:** [More Information Needed]
- **Funded by [optional]:** [More Information Needed]
- **Shared by [optional]:** [More Information Needed]
- **Model type:** [More Information Needed]
- **Language(s) (NLP):** [More Information Needed]
- **License:** [More Information Needed]
- **Finetuned from model [optional]:** [More Information Needed]
### Model Sources [optional]
<!-- Provide the basic links for the model. -->
- **Repository:** [More Information Needed]
- **Paper [optional]:** [More Information Needed]
- **Demo [optional]:** [More Information Needed]
## Uses
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
### Direct Use
<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
[More Information Needed]
### Downstream Use [optional]
<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
[More Information Needed]
### Out-of-Scope Use
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
[More Information Needed]
## Bias, Risks, and Limitations
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
[More Information Needed]
### Recommendations
<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
## How to Get Started with the Model
Use the code below to get started with the model.
[More Information Needed]
## Training Details
### Training Data
<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
[More Information Needed]
### Training Procedure
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
#### Preprocessing [optional]
[More Information Needed]
#### Training Hyperparameters
- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
#### Speeds, Sizes, Times [optional]
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
[More Information Needed]
## Evaluation
<!-- This section describes the evaluation protocols and provides the results. -->
### Testing Data, Factors & Metrics
#### Testing Data
<!-- This should link to a Dataset Card if possible. -->
[More Information Needed]
#### Factors
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
[More Information Needed]
#### Metrics
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
[More Information Needed]
### Results
[More Information Needed]
#### Summary
## Model Examination [optional]
<!-- Relevant interpretability work for the model goes here -->
[More Information Needed]
## Environmental Impact
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
- **Hardware Type:** [More Information Needed]
- **Hours used:** [More Information Needed]
- **Cloud Provider:** [More Information Needed]
- **Compute Region:** [More Information Needed]
- **Carbon Emitted:** [More Information Needed]
## Technical Specifications [optional]
### Model Architecture and Objective
[More Information Needed]
### Compute Infrastructure
[More Information Needed]
#### Hardware
[More Information Needed]
#### Software
[More Information Needed]
## Citation [optional]
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
**BibTeX:**
[More Information Needed]
**APA:**
[More Information Needed]
## Glossary [optional]
<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
[More Information Needed]
## More Information [optional]
[More Information Needed]
## Model Card Authors [optional]
[More Information Needed]
## Model Card Contact
[More Information Needed]
|
Praha-Labs/LFM-NEW
|
Praha-Labs
| 2025-08-29T18:21:27Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"lfm2",
"text-generation",
"text-generation-inference",
"unsloth",
"trl",
"conversational",
"en",
"base_model:Vyvo/VyvoTTS-LFM2-Neuvillette",
"base_model:finetune:Vyvo/VyvoTTS-LFM2-Neuvillette",
"license:apache-2.0",
"autotrain_compatible",
"endpoints_compatible",
"region:us"
] |
text-generation
| 2025-08-29T18:21:10Z |
---
base_model: Vyvo/VyvoTTS-LFM2-Neuvillette
tags:
- text-generation-inference
- transformers
- unsloth
- lfm2
- trl
license: apache-2.0
language:
- en
---
# Uploaded model
- **Developed by:** Praha-Labs
- **License:** apache-2.0
- **Finetuned from model :** Vyvo/VyvoTTS-LFM2-Neuvillette
This lfm2 model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
|
goptouy/blockassist-bc-secretive_unseen_python_1756491661
|
goptouy
| 2025-08-29T18:21:19Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"secretive unseen python",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:21:02Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- secretive unseen python
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
goptouy/blockassist-bc-colorful_smooth_elk_1756491535
|
goptouy
| 2025-08-29T18:19:14Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"colorful smooth elk",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:18:56Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- colorful smooth elk
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
dsbgfdqew4/FULL.Video.18.do.surfista.vazado.video.do.surfista.no.banheiro.surfista.mansao.privilege.erome
|
dsbgfdqew4
| 2025-08-29T18:19:10Z | 0 | 0 | null |
[
"region:us"
] | null | 2025-08-29T18:18:28Z |
<a href="https://allyoutubers.com/Portal-Zacarias-Vídeo"> 🌐 FULL.Video.18.do.surfista.vazado.video.do.surfista.no.banheiro.surfista.mansao.privilege.erome
🔴 ➤►DOWNLOAD👉👉🟢 ➤ <a href="https://allyoutubers.com/Portal-Zacarias-Vídeo"> 🌐 FULL.Video.18.do.surfista.vazado.video.do.surfista.no.banheiro.surfista.mansao.privilege.erome
<a href="https://allyoutubers.com/Portal-Zacarias-Vídeo"> 🌐 FULL.Video.18.do.surfista.vazado.video.do.surfista.no.banheiro.surfista.mansao.privilege.erome
🔴 ➤►DOWNLOAD👉👉🟢 ➤ <a href="https://allyoutubers.com/Portal-Zacarias-Vídeo"> 🌐 FULL.Video.18.do.surfista.vazado.video.do.surfista.no.banheiro.surfista.mansao.privilege.erome
|
DannyAI/Fine_Tune_GPRO_Mistral_v0.3_7B
|
DannyAI
| 2025-08-29T18:19:01Z | 0 | 0 | null |
[
"safetensors",
"unsloth",
"license:mit",
"region:us"
] | null | 2025-08-29T15:43:48Z |
---
license: mit
tags:
- unsloth
---
|
mradermacher/XiYanSQL-QwenCoder-14B-2504-i1-GGUF
|
mradermacher
| 2025-08-29T18:18:19Z | 0 | 0 | null |
[
"gguf",
"endpoints_compatible",
"region:us",
"imatrix",
"conversational"
] | null | 2025-08-29T16:59:43Z |
<!-- ### quantize_version: 2 -->
<!-- ### output_tensor_quantised: 1 -->
<!-- ### convert_type: hf -->
<!-- ### vocab_type: -->
<!-- ### tags: nicoboss -->
<!-- ### quants: Q2_K IQ3_M Q4_K_S IQ3_XXS Q3_K_M small-IQ4_NL Q4_K_M IQ2_M Q6_K IQ4_XS Q2_K_S IQ1_M Q3_K_S IQ2_XXS Q3_K_L IQ2_XS Q5_K_S IQ2_S IQ1_S Q5_K_M Q4_0 IQ3_XS Q4_1 IQ3_S -->
<!-- ### quants_skip: -->
<!-- ### skip_mmproj: -->
weighted/imatrix quants of https://huggingface.co/XGenerationLab/XiYanSQL-QwenCoder-14B-2504
|
goptouy/blockassist-bc-secretive_unseen_python_1756491440
|
goptouy
| 2025-08-29T18:17:39Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"secretive unseen python",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:17:21Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- secretive unseen python
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
AnonymousCS/populism_classifier_bsample_188
|
AnonymousCS
| 2025-08-29T18:16:32Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"xlm-roberta",
"text-classification",
"generated_from_trainer",
"base_model:AnonymousCS/populism_xlmr_base",
"base_model:finetune:AnonymousCS/populism_xlmr_base",
"license:mit",
"autotrain_compatible",
"endpoints_compatible",
"region:us"
] |
text-classification
| 2025-08-29T18:14:17Z |
---
library_name: transformers
license: mit
base_model: AnonymousCS/populism_xlmr_base
tags:
- generated_from_trainer
metrics:
- accuracy
model-index:
- name: populism_classifier_bsample_188
results: []
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# populism_classifier_bsample_188
This model is a fine-tuned version of [AnonymousCS/populism_xlmr_base](https://huggingface.co/AnonymousCS/populism_xlmr_base) on the None dataset.
It achieves the following results on the evaluation set:
- Loss: 0.7677
- Accuracy: 0.0625
- 1-f1: 0.1176
- 1-recall: 1.0
- 1-precision: 0.0625
- Balanced Acc: 0.5
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-06
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.06
- num_epochs: 15
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy | 1-f1 | 1-recall | 1-precision | Balanced Acc |
|:-------------:|:-----:|:----:|:---------------:|:--------:|:------:|:--------:|:-----------:|:------------:|
| 0.7395 | 1.0 | 15 | 0.7633 | 0.0625 | 0.1176 | 1.0 | 0.0625 | 0.5 |
| 0.6978 | 2.0 | 30 | 0.7616 | 0.0625 | 0.1176 | 1.0 | 0.0625 | 0.5 |
| 0.7113 | 3.0 | 45 | 0.7629 | 0.0625 | 0.1176 | 1.0 | 0.0625 | 0.5 |
| 0.705 | 4.0 | 60 | 0.7651 | 0.0625 | 0.1176 | 1.0 | 0.0625 | 0.5 |
| 0.7479 | 5.0 | 75 | 0.7672 | 0.0625 | 0.1176 | 1.0 | 0.0625 | 0.5 |
| 0.6759 | 6.0 | 90 | 0.7664 | 0.0625 | 0.1176 | 1.0 | 0.0625 | 0.5 |
| 0.6576 | 7.0 | 105 | 0.7677 | 0.0625 | 0.1176 | 1.0 | 0.0625 | 0.5 |
### Framework versions
- Transformers 4.46.3
- Pytorch 2.4.1+cu121
- Datasets 3.1.0
- Tokenizers 0.20.3
|
goptouy/blockassist-bc-carnivorous_tawny_stingray_1756491344
|
goptouy
| 2025-08-29T18:16:02Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"carnivorous tawny stingray",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:15:45Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- carnivorous tawny stingray
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
alok0777/blockassist-bc-masked_pensive_lemur_1756491287
|
alok0777
| 2025-08-29T18:15:59Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"masked pensive lemur",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:15:33Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- masked pensive lemur
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
wsbagnsv1/InternVL3_5-38B-Q5_K_S-GGUF
|
wsbagnsv1
| 2025-08-29T18:14:46Z | 0 | 0 |
transformers
|
[
"transformers",
"gguf",
"internvl",
"custom_code",
"llama-cpp",
"gguf-my-repo",
"image-text-to-text",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"base_model:OpenGVLab/InternVL3_5-38B",
"base_model:finetune:OpenGVLab/InternVL3_5-38B",
"license:apache-2.0",
"endpoints_compatible",
"region:us"
] |
image-text-to-text
| 2025-08-29T18:13:20Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model: OpenGVLab/InternVL3_5-38B
base_model_relation: finetune
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
- llama-cpp
- gguf-my-repo
---
# wsbagnsv1/InternVL3_5-38B-Q5_K_S-GGUF
This model was converted to GGUF format from [`OpenGVLab/InternVL3_5-38B`](https://huggingface.co/OpenGVLab/InternVL3_5-38B) using llama.cpp via the ggml.ai's [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space.
Refer to the [original model card](https://huggingface.co/OpenGVLab/InternVL3_5-38B) for more details on the model.
## Use with llama.cpp
Install llama.cpp through brew (works on Mac and Linux)
```bash
brew install llama.cpp
```
Invoke the llama.cpp server or the CLI.
### CLI:
```bash
llama-cli --hf-repo wsbagnsv1/InternVL3_5-38B-Q5_K_S-GGUF --hf-file internvl3_5-38b-q5_k_s.gguf -p "The meaning to life and the universe is"
```
### Server:
```bash
llama-server --hf-repo wsbagnsv1/InternVL3_5-38B-Q5_K_S-GGUF --hf-file internvl3_5-38b-q5_k_s.gguf -c 2048
```
Note: You can also use this checkpoint directly through the [usage steps](https://github.com/ggerganov/llama.cpp?tab=readme-ov-file#usage) listed in the Llama.cpp repo as well.
Step 1: Clone llama.cpp from GitHub.
```
git clone https://github.com/ggerganov/llama.cpp
```
Step 2: Move into the llama.cpp folder and build it with `LLAMA_CURL=1` flag along with other hardware-specific flags (for ex: LLAMA_CUDA=1 for Nvidia GPUs on Linux).
```
cd llama.cpp && LLAMA_CURL=1 make
```
Step 3: Run inference through the main binary.
```
./llama-cli --hf-repo wsbagnsv1/InternVL3_5-38B-Q5_K_S-GGUF --hf-file internvl3_5-38b-q5_k_s.gguf -p "The meaning to life and the universe is"
```
or
```
./llama-server --hf-repo wsbagnsv1/InternVL3_5-38B-Q5_K_S-GGUF --hf-file internvl3_5-38b-q5_k_s.gguf -c 2048
```
|
AnonymousCS/populism_classifier_bsample_187
|
AnonymousCS
| 2025-08-29T18:13:53Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"xlm-roberta",
"text-classification",
"generated_from_trainer",
"base_model:AnonymousCS/populism_xlmr_base",
"base_model:finetune:AnonymousCS/populism_xlmr_base",
"license:mit",
"autotrain_compatible",
"endpoints_compatible",
"region:us"
] |
text-classification
| 2025-08-29T18:11:05Z |
---
library_name: transformers
license: mit
base_model: AnonymousCS/populism_xlmr_base
tags:
- generated_from_trainer
metrics:
- accuracy
model-index:
- name: populism_classifier_bsample_187
results: []
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# populism_classifier_bsample_187
This model is a fine-tuned version of [AnonymousCS/populism_xlmr_base](https://huggingface.co/AnonymousCS/populism_xlmr_base) on the None dataset.
It achieves the following results on the evaluation set:
- Loss: 0.7374
- Accuracy: 0.0359
- 1-f1: 0.0693
- 1-recall: 1.0
- 1-precision: 0.0359
- Balanced Acc: 0.5
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-06
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.06
- num_epochs: 15
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy | 1-f1 | 1-recall | 1-precision | Balanced Acc |
|:-------------:|:-----:|:----:|:---------------:|:--------:|:------:|:--------:|:-----------:|:------------:|
| 0.7144 | 1.0 | 13 | 0.7651 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.7109 | 2.0 | 26 | 0.7570 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.7222 | 3.0 | 39 | 0.7542 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.6849 | 4.0 | 52 | 0.7506 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.7012 | 5.0 | 65 | 0.7478 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.7035 | 6.0 | 78 | 0.7465 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.6665 | 7.0 | 91 | 0.7447 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.7684 | 8.0 | 104 | 0.7451 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.7179 | 9.0 | 117 | 0.7447 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.6673 | 10.0 | 130 | 0.7428 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.7015 | 11.0 | 143 | 0.7410 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.6969 | 12.0 | 156 | 0.7392 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.7305 | 13.0 | 169 | 0.7378 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.699 | 14.0 | 182 | 0.7374 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
| 0.7246 | 15.0 | 195 | 0.7374 | 0.0359 | 0.0693 | 1.0 | 0.0359 | 0.5 |
### Framework versions
- Transformers 4.46.3
- Pytorch 2.4.1+cu121
- Datasets 3.1.0
- Tokenizers 0.20.3
|
sekirr/blockassist-bc-masked_tenacious_whale_1756491155
|
sekirr
| 2025-08-29T18:13:14Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"masked tenacious whale",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:13:11Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- masked tenacious whale
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
shra233/mindspace
|
shra233
| 2025-08-29T18:11:50Z | 0 | 0 |
peft
|
[
"peft",
"safetensors",
"base_model:adapter:meta-llama/Llama-2-7b-chat-hf",
"lora",
"sft",
"trl",
"text-generation",
"conversational",
"arxiv:1910.09700",
"base_model:meta-llama/Llama-2-7b-chat-hf",
"region:us"
] |
text-generation
| 2025-08-29T13:29:48Z |
---
base_model: meta-llama/Llama-2-7b-chat-hf
library_name: peft
pipeline_tag: text-generation
tags:
- base_model:adapter:meta-llama/Llama-2-7b-chat-hf
- lora
- sft
- trl
---
# Model Card for Model ID
<!-- Provide a quick summary of what the model is/does. -->
## Model Details
### Model Description
<!-- Provide a longer summary of what this model is. -->
- **Developed by:** [More Information Needed]
- **Funded by [optional]:** [More Information Needed]
- **Shared by [optional]:** [More Information Needed]
- **Model type:** [More Information Needed]
- **Language(s) (NLP):** [More Information Needed]
- **License:** [More Information Needed]
- **Finetuned from model [optional]:** [More Information Needed]
### Model Sources [optional]
<!-- Provide the basic links for the model. -->
- **Repository:** [More Information Needed]
- **Paper [optional]:** [More Information Needed]
- **Demo [optional]:** [More Information Needed]
## Uses
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
### Direct Use
<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
[More Information Needed]
### Downstream Use [optional]
<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
[More Information Needed]
### Out-of-Scope Use
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
[More Information Needed]
## Bias, Risks, and Limitations
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
[More Information Needed]
### Recommendations
<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
## How to Get Started with the Model
Use the code below to get started with the model.
[More Information Needed]
## Training Details
### Training Data
<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
[More Information Needed]
### Training Procedure
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
#### Preprocessing [optional]
[More Information Needed]
#### Training Hyperparameters
- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
#### Speeds, Sizes, Times [optional]
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
[More Information Needed]
## Evaluation
<!-- This section describes the evaluation protocols and provides the results. -->
### Testing Data, Factors & Metrics
#### Testing Data
<!-- This should link to a Dataset Card if possible. -->
[More Information Needed]
#### Factors
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
[More Information Needed]
#### Metrics
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
[More Information Needed]
### Results
[More Information Needed]
#### Summary
## Model Examination [optional]
<!-- Relevant interpretability work for the model goes here -->
[More Information Needed]
## Environmental Impact
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
- **Hardware Type:** [More Information Needed]
- **Hours used:** [More Information Needed]
- **Cloud Provider:** [More Information Needed]
- **Compute Region:** [More Information Needed]
- **Carbon Emitted:** [More Information Needed]
## Technical Specifications [optional]
### Model Architecture and Objective
[More Information Needed]
### Compute Infrastructure
[More Information Needed]
#### Hardware
[More Information Needed]
#### Software
[More Information Needed]
## Citation [optional]
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
**BibTeX:**
[More Information Needed]
**APA:**
[More Information Needed]
## Glossary [optional]
<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
[More Information Needed]
## More Information [optional]
[More Information Needed]
## Model Card Authors [optional]
[More Information Needed]
## Model Card Contact
[More Information Needed]
### Framework versions
- PEFT 0.17.1
|
goptouy/blockassist-bc-colorful_smooth_elk_1756491032
|
goptouy
| 2025-08-29T18:10:50Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"colorful smooth elk",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:10:33Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- colorful smooth elk
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
AnonymousCS/populism_classifier_bsample_186
|
AnonymousCS
| 2025-08-29T18:10:41Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"xlm-roberta",
"text-classification",
"generated_from_trainer",
"base_model:AnonymousCS/populism_xlmr_base",
"base_model:finetune:AnonymousCS/populism_xlmr_base",
"license:mit",
"autotrain_compatible",
"endpoints_compatible",
"region:us"
] |
text-classification
| 2025-08-29T18:07:05Z |
---
library_name: transformers
license: mit
base_model: AnonymousCS/populism_xlmr_base
tags:
- generated_from_trainer
metrics:
- accuracy
model-index:
- name: populism_classifier_bsample_186
results: []
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# populism_classifier_bsample_186
This model is a fine-tuned version of [AnonymousCS/populism_xlmr_base](https://huggingface.co/AnonymousCS/populism_xlmr_base) on the None dataset.
It achieves the following results on the evaluation set:
- Loss: 0.5681
- Accuracy: 0.9801
- 1-f1: 0.0
- 1-recall: 0.0
- 1-precision: 0.0
- Balanced Acc: 0.5
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-06
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.06
- num_epochs: 15
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy | 1-f1 | 1-recall | 1-precision | Balanced Acc |
|:-------------:|:-----:|:----:|:---------------:|:--------:|:------:|:--------:|:-----------:|:------------:|
| 0.8574 | 1.0 | 29 | 1.0432 | 0.0199 | 0.0390 | 1.0 | 0.0199 | 0.5 |
| 0.7502 | 2.0 | 58 | 0.9557 | 0.0199 | 0.0390 | 1.0 | 0.0199 | 0.5 |
| 0.7956 | 3.0 | 87 | 0.8820 | 0.0199 | 0.0390 | 1.0 | 0.0199 | 0.5 |
| 0.7847 | 4.0 | 116 | 0.8200 | 0.0199 | 0.0390 | 1.0 | 0.0199 | 0.5 |
| 0.7748 | 5.0 | 145 | 0.7680 | 0.0199 | 0.0390 | 1.0 | 0.0199 | 0.5 |
| 0.7467 | 6.0 | 174 | 0.7224 | 0.0199 | 0.0390 | 1.0 | 0.0199 | 0.5 |
| 0.7262 | 7.0 | 203 | 0.6816 | 0.9801 | 0.0 | 0.0 | 0.0 | 0.5 |
| 0.654 | 8.0 | 232 | 0.6507 | 0.9801 | 0.0 | 0.0 | 0.0 | 0.5 |
| 0.6418 | 9.0 | 261 | 0.6255 | 0.9801 | 0.0 | 0.0 | 0.0 | 0.5 |
| 0.5767 | 10.0 | 290 | 0.6068 | 0.9801 | 0.0 | 0.0 | 0.0 | 0.5 |
| 0.5875 | 11.0 | 319 | 0.5928 | 0.9801 | 0.0 | 0.0 | 0.0 | 0.5 |
| 0.6796 | 12.0 | 348 | 0.5830 | 0.9801 | 0.0 | 0.0 | 0.0 | 0.5 |
| 0.6369 | 13.0 | 377 | 0.5755 | 0.9801 | 0.0 | 0.0 | 0.0 | 0.5 |
| 0.5774 | 14.0 | 406 | 0.5704 | 0.9801 | 0.0 | 0.0 | 0.0 | 0.5 |
| 0.6093 | 15.0 | 435 | 0.5681 | 0.9801 | 0.0 | 0.0 | 0.0 | 0.5 |
### Framework versions
- Transformers 4.46.3
- Pytorch 2.4.1+cu121
- Datasets 3.1.0
- Tokenizers 0.20.3
|
goptouy/blockassist-bc-carnivorous_tawny_stingray_1756491000
|
goptouy
| 2025-08-29T18:10:19Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"carnivorous tawny stingray",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:10:01Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- carnivorous tawny stingray
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
goptouy/blockassist-bc-wary_lanky_porcupine_1756490904
|
goptouy
| 2025-08-29T18:08:42Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"wary lanky porcupine",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:08:25Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- wary lanky porcupine
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
goptouy/blockassist-bc-stinky_tricky_swan_1756490843
|
goptouy
| 2025-08-29T18:07:41Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"stinky tricky swan",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:07:24Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- stinky tricky swan
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
acidjp/blockassist-bc-pesty_extinct_prawn_1756488543
|
acidjp
| 2025-08-29T18:07:14Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"pesty extinct prawn",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:07:10Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- pesty extinct prawn
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
youryoui/blockassist-bc-beaked_frisky_ox_1756490692
|
youryoui
| 2025-08-29T18:05:04Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"beaked frisky ox",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:04:52Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- beaked frisky ox
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
seraphimzzzz/547876
|
seraphimzzzz
| 2025-08-29T18:05:02Z | 0 | 0 | null |
[
"region:us"
] | null | 2025-08-29T18:04:59Z |
[View on Civ Archive](https://civarchive.com/models/568173?modelVersionId=633199)
|
ACampero/exp2_standard_em_1B_new
|
ACampero
| 2025-08-29T18:04:54Z | 0 | 0 |
transformers
|
[
"transformers",
"safetensors",
"unsloth",
"arxiv:1910.09700",
"endpoints_compatible",
"region:us"
] | null | 2025-08-29T17:23:43Z |
---
library_name: transformers
tags:
- unsloth
---
# Model Card for Model ID
<!-- Provide a quick summary of what the model is/does. -->
## Model Details
### Model Description
<!-- Provide a longer summary of what this model is. -->
This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
- **Developed by:** [More Information Needed]
- **Funded by [optional]:** [More Information Needed]
- **Shared by [optional]:** [More Information Needed]
- **Model type:** [More Information Needed]
- **Language(s) (NLP):** [More Information Needed]
- **License:** [More Information Needed]
- **Finetuned from model [optional]:** [More Information Needed]
### Model Sources [optional]
<!-- Provide the basic links for the model. -->
- **Repository:** [More Information Needed]
- **Paper [optional]:** [More Information Needed]
- **Demo [optional]:** [More Information Needed]
## Uses
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
### Direct Use
<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
[More Information Needed]
### Downstream Use [optional]
<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
[More Information Needed]
### Out-of-Scope Use
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
[More Information Needed]
## Bias, Risks, and Limitations
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
[More Information Needed]
### Recommendations
<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
## How to Get Started with the Model
Use the code below to get started with the model.
[More Information Needed]
## Training Details
### Training Data
<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
[More Information Needed]
### Training Procedure
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
#### Preprocessing [optional]
[More Information Needed]
#### Training Hyperparameters
- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
#### Speeds, Sizes, Times [optional]
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
[More Information Needed]
## Evaluation
<!-- This section describes the evaluation protocols and provides the results. -->
### Testing Data, Factors & Metrics
#### Testing Data
<!-- This should link to a Dataset Card if possible. -->
[More Information Needed]
#### Factors
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
[More Information Needed]
#### Metrics
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
[More Information Needed]
### Results
[More Information Needed]
#### Summary
## Model Examination [optional]
<!-- Relevant interpretability work for the model goes here -->
[More Information Needed]
## Environmental Impact
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
- **Hardware Type:** [More Information Needed]
- **Hours used:** [More Information Needed]
- **Cloud Provider:** [More Information Needed]
- **Compute Region:** [More Information Needed]
- **Carbon Emitted:** [More Information Needed]
## Technical Specifications [optional]
### Model Architecture and Objective
[More Information Needed]
### Compute Infrastructure
[More Information Needed]
#### Hardware
[More Information Needed]
#### Software
[More Information Needed]
## Citation [optional]
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
**BibTeX:**
[More Information Needed]
**APA:**
[More Information Needed]
## Glossary [optional]
<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
[More Information Needed]
## More Information [optional]
[More Information Needed]
## Model Card Authors [optional]
[More Information Needed]
## Model Card Contact
[More Information Needed]
|
seraphimzzzz/550138
|
seraphimzzzz
| 2025-08-29T18:04:38Z | 0 | 0 | null |
[
"region:us"
] | null | 2025-08-29T18:04:35Z |
[View on Civ Archive](https://civarchive.com/models/568173?modelVersionId=635426)
|
goptouy/blockassist-bc-omnivorous_soaring_pigeon_1756490643
|
goptouy
| 2025-08-29T18:04:21Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"omnivorous soaring pigeon",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:04:04Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- omnivorous soaring pigeon
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
liukevin666/blockassist-bc-yawning_striped_cassowary_1756490546
|
liukevin666
| 2025-08-29T18:03:41Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"yawning striped cassowary",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:03:21Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- yawning striped cassowary
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
goptouy/blockassist-bc-stinky_tricky_swan_1756490549
|
goptouy
| 2025-08-29T18:02:47Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"stinky tricky swan",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T18:02:30Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- stinky tricky swan
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
goptouy/blockassist-bc-carnivorous_tawny_stingray_1756490392
|
goptouy
| 2025-08-29T18:00:11Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"carnivorous tawny stingray",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T17:59:52Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- carnivorous tawny stingray
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview
|
OpenGVLab
| 2025-08-29T17:59:02Z | 3,599 | 46 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternViT-300M-448px-V2_5",
"base_model:merge:OpenGVLab/InternViT-300M-448px-V2_5",
"base_model:openai/gpt-oss-20b",
"base_model:merge:openai/gpt-oss-20b",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:37Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternViT-300M-448px-V2_5
- openai/gpt-oss-20b
base_model_relation: merge
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-GPT-OSS-20B-A4B-Preview
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### vLLM
vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs and MLLMs.
Please refer to the [documentation](https://docs.vllm.ai/en/latest/examples/offline_inference/vision_language.html?h=internvl#vision-language) for how to deploy internvl series.
```sh
pip install vllm>=0.10.1
```
NOTE: Up to version 0.10.1.1, vLLM exhibits compatibility issues with GPT-OSS when applied in MLLMs. If you encounter any errors, please try replacing the `vllm/model_executor/models/gpt_oss.py` file with the following content:
```python
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Iterable
from typing import Optional
import torch
import torch.distributed as dist
from torch import nn
from transformers import GptOssConfig
from vllm.attention import Attention, AttentionType
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (get_ep_group, get_tensor_model_parallel_rank, get_pp_group,
get_tensor_model_parallel_world_size)
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors
from vllm.utils import cdiv
from .utils import (extract_layer_index,
make_empty_intermediate_tensors_factory,
maybe_prefix)
class OAIAttention(nn.Module):
def __init__(
self,
config: GptOssConfig,
quant_config: Optional[QuantizationConfig] = None,
cache_config: Optional[CacheConfig] = None,
prefix: str = "",
):
super().__init__()
self.layer_idx = extract_layer_index(prefix)
self.head_dim = config.head_dim
self.num_attention_heads = config.num_attention_heads
self.num_key_value_heads = config.num_key_value_heads
self.hidden_size = config.hidden_size
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=config.max_position_embeddings,
base=config.rope_theta,
dtype=torch.float32,
rope_scaling={
"rope_type":
"yarn",
"factor":
config.rope_scaling["factor"],
"original_max_position_embeddings":
config.rope_scaling["original_max_position_embeddings"],
"beta_fast":
config.rope_scaling["beta_fast"],
"beta_slow":
config.rope_scaling["beta_slow"],
},
is_neox_style=True,
)
tp_size = get_tensor_model_parallel_world_size()
self.sinks = torch.nn.Parameter(
torch.empty(config.num_attention_heads // tp_size,
dtype=torch.bfloat16,
requires_grad=False))
self.norm = RMSNorm(config.hidden_size, eps=1e-5)
self.q_size = self.num_attention_heads * self.head_dim // tp_size
self.kv_size = self.num_key_value_heads * self.head_dim // tp_size
self.scaling = self.head_dim**-0.5
self.rope_theta = config.rope_theta
self.qkv = QKVParallelLinear(
hidden_size=self.hidden_size,
head_size=self.head_dim,
total_num_heads=self.num_attention_heads,
total_num_kv_heads=self.num_key_value_heads,
quant_config=quant_config,
prefix=f"{prefix}.qkv_proj",
)
self.o_proj = RowParallelLinear(
input_size=self.num_attention_heads * self.head_dim,
output_size=self.hidden_size,
quant_config=quant_config,
prefix=f"{prefix}.o_proj",
)
self.num_local_attention_heads = config.num_attention_heads // tp_size
self.num_local_key_value_heads = config.num_key_value_heads // tp_size
# Only apply sliding window to every other layer
sliding_window = (config.sliding_window if self.layer_idx %
2 == 0 else None)
self.attn = Attention(
self.num_local_attention_heads,
self.head_dim,
self.scaling,
num_kv_heads=self.num_local_key_value_heads,
cache_config=cache_config,
quant_config=quant_config,
per_layer_sliding_window=sliding_window,
attn_type=AttentionType.DECODER,
prefix=f"{prefix}.attn",
sinks=self.sinks,
)
def forward(self, hidden_states: torch.Tensor,
positions: torch.Tensor) -> torch.Tensor:
t = self.norm(hidden_states)
qkv, _ = self.qkv(t)
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
q, k = self.rotary_emb(positions, q, k)
v = v.contiguous()
attn_output = self.attn(q, k, v)
output, _ = self.o_proj(attn_output)
return output + hidden_states
class MLPBlock(torch.nn.Module):
def __init__(
self,
config: GptOssConfig,
layer_idx: int,
quant_config: QuantizationConfig,
prefix: str = "",
):
super().__init__()
self.layer_idx = layer_idx
self.num_experts = config.num_local_experts
self.experts_per_token = config.num_experts_per_tok
# self.world_size = dist.get_world_size() if dist.is_initialized() else 1
self.norm = RMSNorm(config.hidden_size, eps=1e-5)
self.router = torch.nn.Linear(config.hidden_size,
config.num_local_experts,
dtype=torch.bfloat16)
# assert config.intermediate_size % self.world_size == 0
self.experts = FusedMoE(num_experts=config.num_local_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.intermediate_size,
reduce_results=True,
renormalize=True,
quant_config=quant_config,
prefix=f"{prefix}.experts",
apply_router_weight_on_input=False,
has_bias=True,
activation="swigluoai")
def forward(self, x: torch.Tensor) -> torch.Tensor:
t = self.norm(x)
g = self.router(t)
t = self.experts(hidden_states=t, router_logits=g)
return x + t
class TransformerBlock(torch.nn.Module):
def __init__(
self,
config: GptOssConfig,
quant_config: QuantizationConfig,
prefix: str = "",
):
super().__init__()
self.layer_idx = extract_layer_index(prefix)
self.attn = OAIAttention(config, prefix=f"{prefix}.attn")
self.mlp = MLPBlock(config,
self.layer_idx,
quant_config=quant_config,
prefix=f"{prefix}.mlp")
def forward(self, hidden_states: torch.Tensor,
positions: torch.Tensor) -> torch.Tensor:
attn_output = self.attn(hidden_states, positions)
output = self.mlp(attn_output)
return output
@support_torch_compile
class GptOssModel(nn.Module):
def __init__(
self,
*,
vllm_config: VllmConfig,
prefix: str = "",
):
super().__init__()
self.config = vllm_config.model_config.hf_config
self.quant_config = vllm_config.quant_config
self.config.hidden_size = self.config.hidden_size
self.embedding = VocabParallelEmbedding(
self.config.vocab_size,
self.config.hidden_size,
)
self.layers = torch.nn.ModuleList([
TransformerBlock(
self.config,
quant_config=self.quant_config,
prefix=maybe_prefix(prefix, f"block.{layer_idx}"),
) for layer_idx in range(self.config.num_hidden_layers)
])
self.norm = RMSNorm(self.config.hidden_size, eps=1e-5)
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], self.config.hidden_size))
def forward(self, input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,) -> torch.Tensor:
if get_pp_group().is_first_rank:
if inputs_embeds is not None:
hidden_states = inputs_embeds
else:
# hidden_states = self.get_input_embeddings(input_ids)
hidden_states = self.embedding(input_ids)
residual = None
else:
assert intermediate_tensors is not None
hidden_states = intermediate_tensors["hidden_states"]
residual = intermediate_tensors["residual"]
# x = self.embedding(input_ids)
# for layer in self.layers:
# x = layer(x, positions)
# x = self.norm(x)
for layer in self.layers:
hidden_states = layer(hidden_states, positions)
hidden_states = self.norm(hidden_states)
return hidden_states
class GptOssForCausalLM(nn.Module):
def __init__(
self,
vllm_config: VllmConfig,
prefix: str = "",
):
super().__init__()
self.vllm_config = vllm_config
self.model_config = vllm_config.model_config.hf_config
self.model = GptOssModel(
vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"),
)
self.lm_head = ParallelLMHead(
self.model_config.vocab_size,
self.model_config.hidden_size,
)
self.logits_processor = LogitsProcessor(self.model_config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
def forward(self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None) -> torch.Tensor:
assert intermediate_tensors is None
assert inputs_embeds is None
return self.model(input_ids, positions)
def compute_logits(self, hidden_states: torch.Tensor,
sampling_metadata: SamplingMetadata) -> torch.Tensor:
logits = self.logits_processor(self.lm_head, hidden_states,
sampling_metadata)
return logits
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.embedding(input_ids)
def _load_weights_mxfp4(
self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
rename_mapping = {
"self_attn": "attn",
"input_layernorm.weight": "attn.norm.weight",
"post_attention_layernorm.weight": "mlp.norm.weight",
"embed_tokens": "embedding",
}
def maybe_rename(name: str) -> str:
for remap_name, new_name in rename_mapping.items():
if remap_name in name:
return name.replace(remap_name, new_name)
return name
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
mxfp4_block = 32
tp_rank = get_tensor_model_parallel_rank()
tp_size = get_tensor_model_parallel_world_size()
intermediate_size = self.model_config.intermediate_size
intermediate_size_block = intermediate_size // mxfp4_block
per_rank_intermediate_size_block = cdiv(intermediate_size_block,
tp_size)
per_rank_intermediate_size = (per_rank_intermediate_size_block *
mxfp4_block)
# Calculate common slicing bounds for current rank
tp_rank_start = tp_rank * per_rank_intermediate_size
tp_rank_end = min((tp_rank + 1) * per_rank_intermediate_size,
intermediate_size)
# Attention heads per rank
heads_per_rank = self.model_config.num_attention_heads // tp_size
head_start = tp_rank * heads_per_rank
use_ep = self.vllm_config.parallel_config.enable_expert_parallel
ep_size = get_ep_group().world_size
ep_rank = get_ep_group().rank
num_experts = self.model_config.num_local_experts
experts_per_rank = num_experts // ep_size
ep_rank_start = ep_rank * experts_per_rank
ep_rank_end = (ep_rank + 1) * experts_per_rank
for name, weight in weights:
# FIXME(woosuk): Remove this after testing.
weight = weight.cuda()
if "gate_up_proj_blocks" in name:
# Handle MLP gate and up projection weights
new_name = name.replace("gate_up_proj_blocks", "w13_weight")
# flat weight from (E, 2 * N, block_size, entry_per_block)
# to (E, 2 * N, -1), shouldn't trigger copy for contiguous
weight = weight.view(num_experts, 2 * intermediate_size,
-1).contiguous()
# Extract gate and up projection parts
# since the weight is shuffled, we can slice directly
if use_ep:
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
else:
narrow_weight = weight[:,
2 * tp_rank_start:2 * tp_rank_end,
...]
param = params_dict[new_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param,
narrow_weight,
weight_name=new_name,
shard_id=None,
expert_id=None)
loaded_params.add(new_name)
elif "down_proj_blocks" in name:
# Handle MLP down projection weights
new_name = name.replace("down_proj_blocks", "w2_weight")
# same flatten here, but since 2 mx4 value are packed in 1
# uint8, divide by 2
weight = weight.view(num_experts, -1,
intermediate_size // 2).contiguous()
if use_ep:
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
else:
narrow_weight = weight[...,
tp_rank_start // 2:tp_rank_end // 2]
param = params_dict[new_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param,
narrow_weight,
weight_name=new_name,
shard_id=None,
expert_id=None)
loaded_params.add(new_name)
elif "gate_up_proj_scales" in name:
# Handle MLP gate and up projection weights scale
new_name = name.replace("gate_up_proj_scales",
"w13_weight_scale")
if use_ep:
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
else:
narrow_weight = weight[:,
2 * tp_rank_start:2 * tp_rank_end,
...]
param = params_dict[new_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param,
narrow_weight,
weight_name=new_name,
shard_id=None,
expert_id=None)
loaded_params.add(new_name)
elif "down_proj_scales" in name:
# Handle MLP down projection weights
new_name = name.replace("down_proj_scales", "w2_weight_scale")
if use_ep:
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
else:
narrow_weight = weight[..., tp_rank_start //
mxfp4_block:tp_rank_end //
mxfp4_block]
param = params_dict[new_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param,
narrow_weight,
weight_name=new_name,
shard_id=None,
expert_id=None)
loaded_params.add(new_name)
elif "gate_up_proj_bias" in name:
# Handle MLP gate and up projection biases
new_name = name.replace("gate_up_proj_bias", "w13_bias")
# Extract gate and up projection bias parts
if use_ep:
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
else:
narrow_weight = weight[:,
2 * tp_rank_start:2 * tp_rank_end]
param = params_dict[new_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param,
narrow_weight,
weight_name=new_name,
shard_id=None,
expert_id=None)
loaded_params.add(new_name)
elif "down_proj_bias" in name:
# Handle MLP down projection bias
new_name = name.replace("down_proj_bias", "w2_bias")
param = params_dict[new_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
if use_ep:
weight = weight[ep_rank_start:ep_rank_end, ...]
else:
# (only load on rank 0 to avoid duplication)
if tp_rank != 0:
weight.zero_()
weight_loader(param,
weight,
weight_name=new_name,
shard_id=None,
expert_id=None)
loaded_params.add(new_name)
elif "sinks" in name:
# Handle attention sinks (distributed across ranks)
name = name.replace("self_attn", "attn")
param = params_dict[name]
narrow_weight = weight.narrow(0, head_start, heads_per_rank)
param.data.copy_(narrow_weight)
loaded_params.add(name)
elif "q_proj" in name or "k_proj" in name or "v_proj" in name:
shard_id = ("q" if "q_proj" in name else
"k" if "k_proj" in name else "v")
name = name.replace("self_attn", "attn")
param_name = name.replace(f"{shard_id}_proj", "qkv")
param = params_dict[param_name]
weight_loader = param.weight_loader
weight_loader(param, weight, loaded_shard_id=shard_id)
loaded_params.add(param_name)
else:
# Handle all other weights with potential renaming
renamed_name = maybe_rename(name)
if renamed_name not in params_dict:
continue
param = params_dict[renamed_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, weight)
loaded_params.add(renamed_name)
return loaded_params
def _load_weights_other(
self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
rename_mapping = {
"self_attn": "attn",
"input_layernorm.weight": "attn.norm.weight",
"post_attention_layernorm.weight": "mlp.norm.weight",
"embed_tokens": "embedding",
}
def maybe_rename(name: str) -> str:
for remap_name, new_name in rename_mapping.items():
if remap_name in name:
return name.replace(remap_name, new_name)
return name
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
tp_rank = get_tensor_model_parallel_rank()
tp_size = get_tensor_model_parallel_world_size()
intermediate_size = self.model_config.intermediate_size
per_rank_intermediate_size = cdiv(intermediate_size, tp_size)
# Calculate common slicing bounds for current rank
tp_rank_start = tp_rank * per_rank_intermediate_size
tp_rank_end = min((tp_rank + 1) * per_rank_intermediate_size,
intermediate_size)
# Attention heads per rank
heads_per_rank = self.model_config.num_attention_heads // tp_size
head_start = tp_rank * heads_per_rank
use_ep = self.vllm_config.parallel_config.enable_expert_parallel
ep_size = get_ep_group().world_size
ep_rank = get_ep_group().rank
num_experts = self.model_config.num_local_experts
experts_per_rank = num_experts // ep_size
ep_rank_start = ep_rank * experts_per_rank
ep_rank_end = (ep_rank + 1) * experts_per_rank
for name, weight in weights:
if ".experts.gate_up_proj" in name and "bias" not in name:
# Handle MLP gate and up projection weights
new_name = name.replace(".experts.gate_up_proj",
".experts.w13_weight")
# Extract gate and up projection parts
# since the weight is shuffled, we can slice directly
if use_ep:
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
else:
narrow_weight = weight[:, :,
2 * tp_rank_start:2 * tp_rank_end]
narrow_weight = narrow_weight.permute(0, 2, 1).contiguous()
param = params_dict[new_name]
param.copy_(narrow_weight)
loaded_params.add(new_name)
elif ".experts.down_proj" in name and "bias" not in name:
# Handle MLP down projection weights
new_name = name.replace(".experts.down_proj",
".experts.w2_weight")
if use_ep:
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
else:
narrow_weight = weight[:, tp_rank_start:tp_rank_end, :]
narrow_weight = narrow_weight.permute(0, 2, 1).contiguous()
param = params_dict[new_name]
param.copy_(narrow_weight)
loaded_params.add(new_name)
elif "gate_up_proj_bias" in name:
# Handle MLP gate and up projection biases
new_name = name.replace("gate_up_proj_bias", "w13_bias")
# Extract gate and up projection bias parts
if use_ep:
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
else:
narrow_weight = weight[:,
2 * tp_rank_start:2 * tp_rank_end]
param = params_dict[new_name]
param.copy_(narrow_weight)
loaded_params.add(new_name)
elif "down_proj_bias" in name:
# Handle MLP down projection bias
new_name = name.replace("down_proj_bias", "w2_bias")
if use_ep:
weight = weight[ep_rank_start:ep_rank_end, ...]
else:
# (only load on rank 0 to avoid duplication)
if tp_rank != 0:
weight.zero_()
param = params_dict[new_name]
param.copy_(weight)
loaded_params.add(new_name)
elif "sinks" in name:
# Handle attention sinks (distributed across ranks)
name = name.replace("self_attn", "attn")
param = params_dict[name]
narrow_weight = weight.narrow(0, head_start, heads_per_rank)
param.data.copy_(narrow_weight)
loaded_params.add(name)
elif "q_proj" in name or "k_proj" in name or "v_proj" in name:
shard_id = ("q" if "q_proj" in name else
"k" if "k_proj" in name else "v")
name = name.replace("self_attn", "attn")
param_name = name.replace(f"{shard_id}_proj", "qkv")
param = params_dict[param_name]
weight_loader = param.weight_loader
weight_loader(param, weight, loaded_shard_id=shard_id)
loaded_params.add(param_name)
else:
# Handle all other weights with potential renaming
renamed_name = maybe_rename(name)
if renamed_name not in params_dict:
continue
param = params_dict[renamed_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, weight)
loaded_params.add(renamed_name)
return loaded_params
def load_weights(self, weights: Iterable[tuple[str,
torch.Tensor]]) -> set[str]:
quant_method = (self.model_config.quantization_config['quant_method']
if hasattr(self.model_config, "quantization_config")
else None)
if quant_method == "mxfp4":
return self._load_weights_mxfp4(weights)
else:
return self._load_weights_other(weights)
```
### LMDeploy
***WARNING: Up to version 0.9.2, lmdeploy does not provide support for GPT-OSS. To deploy InternVL3_5-GPT-OSS-20B-Preview, we recommend using vLLM.***
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
Dr-Wong-Lu-Yang-CCTV-scandal-video/dr.wong.lu.yang.cctv.kelantan.doctor.video.video.links
|
Dr-Wong-Lu-Yang-CCTV-scandal-video
| 2025-08-29T17:58:15Z | 0 | 0 | null |
[
"region:us"
] | null | 2025-08-29T17:58:04Z |
<animated-image data-catalyst=""><a href="https://tinyurl.com/52jc3rtk" rel="nofollow" data-target="animated-image.originalLink"><img src="https://static.wixstatic.com/media/b249f9_adac8f70fb3f45b88691696c77de18f3~mv2.gif" alt="Foo" data-canonical-src="https://static.wixstatic.com/media/b249f9_adac8f70fb3f45b88691696c77de18f3~mv2.gif" style="max-width: 100%; display: inline-block;" data-target="animated-image.originalImage"></a>
|
goptouy/blockassist-bc-colorful_smooth_elk_1756490267
|
goptouy
| 2025-08-29T17:58:05Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"colorful smooth elk",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T17:57:48Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- colorful smooth elk
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
pytorch/Qwen3-8B-INT4
|
pytorch
| 2025-08-29T17:57:35Z | 152 | 2 |
transformers
|
[
"transformers",
"pytorch",
"qwen3",
"text-generation",
"torchao",
"code",
"math",
"chat",
"conversational",
"multilingual",
"arxiv:2507.16099",
"base_model:Qwen/Qwen3-8B",
"base_model:quantized:Qwen/Qwen3-8B",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"endpoints_compatible",
"region:us"
] |
text-generation
| 2025-05-07T02:45:17Z |
---
library_name: transformers
tags:
- torchao
- code
- math
- chat
- conversational
language:
- multilingual
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3-8B/blob/main/LICENSE
pipeline_tag: text-generation
base_model:
- Qwen/Qwen3-8B
---
[Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) quantized with [torchao](https://huggingface.co/docs/transformers/main/en/quantization/torchao) int4 weight only quantization, using [hqq](https://mobiusml.github.io/hqq_blog/) algorithm for improved accuracy, by PyTorch team.
Use it directly or serve using [vLLM](https://docs.vllm.ai/en/latest/) for 62% VRAM reduction (6.27 GB needed) and 1.2x speedup on A100 GPUs.
# Inference with vLLM
Install vllm nightly and torchao nightly to get some recent changes:
```
pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
pip install torchao
```
## Serving
Then we can serve with the following command:
```Shell
# Server
export MODEL=pytorch/Qwen3-8B-INT4
VLLM_DISABLE_COMPILE_CACHE=1 vllm serve $MODEL --tokenizer $MODEL -O3
```
```Shell
# Client
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "pytorch/Qwen3-8B-INT4",
"messages": [
{"role": "user", "content": "Give me a short introduction to large language models."}
],
"temperature": 0.6,
"top_p": 0.95,
"top_k": 20,
"max_tokens": 32768
}'
```
Note: please use `VLLM_DISABLE_COMPILE_CACHE=1` to disable compile cache when running this code, e.g. `VLLM_DISABLE_COMPILE_CACHE=1 python example.py`, since there are some issues with the composability of compile in vLLM and torchao,
this is expected be resolved in pytorch 2.8.
# Inference with Transformers
Install the required packages:
```Shell
pip install git+https://github.com/huggingface/transformers@main
pip install torchao
pip install torch
pip install accelerate
```
Example:
```Py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "pytorch/Qwen3-8B-INT4"
# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)
# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
# conduct text completion
generated_ids = model.generate(
**model_inputs,
max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
# parsing thinking content
try:
# rindex finding 151668 (</think>)
index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
index = 0
thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")
print("thinking content:", thinking_content)
print("content:", content)
```
# Quantization Recipe
Install the required packages:
```Shell
pip install git+https://github.com/huggingface/transformers@main
pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126
pip install torch
pip install accelerate
```
Use the following code to get the quantized model:
```Py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig
model_id = "Qwen/Qwen3-8B"
from torchao.quantization import Int4WeightOnlyConfig
quant_config = Int4WeightOnlyConfig(group_size=128, use_hqq=True)
quantization_config = TorchAoConfig(quant_type=quant_config)
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16, quantization_config=quantization_config)
tokenizer = AutoTokenizer.from_pretrained(model_id)
# Push to hub
USER_ID = "YOUR_USER_ID"
MODEL_NAME = model_id.split("/")[-1]
save_to = f"{USER_ID}/{MODEL_NAME}-INT4"
quantized_model.push_to_hub(save_to, safe_serialization=False)
tokenizer.push_to_hub(save_to)
# Manual Testing
prompt = "Hey, are you conscious? Can you talk to me?"
messages = [
{
"role": "system",
"content": "",
},
{"role": "user", "content": prompt},
]
templated_prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
print("Prompt:", prompt)
print("Templated prompt:", templated_prompt)
inputs = tokenizer(
templated_prompt,
return_tensors="pt",
).to("cuda")
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128)
output_text = tokenizer.batch_decode(
generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Response:", output_text[0][len(prompt):])
```
Note: to `push_to_hub` you need to run
```Shell
pip install -U "huggingface_hub[cli]"
huggingface-cli login
```
and use a token with write access, from https://huggingface.co/settings/tokens
# Model Quality
We rely on [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) to evaluate the quality of the quantized model.
| Benchmark | | |
|----------------------------------|----------------|---------------------------|
| | Qwen3-8B | Qwen3-8B-INT4 |
| **General** | | |
| mmlu | 73.04 | 70.4 |
| mmlu_pro | 53.81 | 52.79 |
| bbh | 79.33 | 74.92 |
| **Multilingual** | | |
| mgsm_en_cot_en | 39.6 | 33.2 |
| m_mmlu (avg) | 57.17 | 54.06 |
| **Math** | | |
| gpqa_main_zeroshot | 35.71 | 32.14 |
| gsm8k | 87.79 | 86.28 |
| leaderboard_math_hard (v3) | 53.7 | 46.83 |
| **Overall** | 60.02 | 56.33 |
<details>
<summary> Reproduce Model Quality Results </summary>
Need to install lm-eval from source:
https://github.com/EleutherAI/lm-evaluation-harness#install
## baseline
```Shell
lm_eval --model hf --model_args pretrained=Qwen/Qwen3-8B --tasks mmlu --device cuda:0 --batch_size 8
```
## int4 weight only quantization with hqq (INT4)
```Shell
export MODEL=pytorch/Qwen3-8B-INT4
# or
# export MODEL=Qwen/Qwen3-8B
lm_eval --model hf --model_args pretrained=$MODEL --tasks mmlu --device cuda:0 --batch_size 8
```
</details>
# Peak Memory Usage
## Results
| Benchmark | | |
|------------------|----------------|--------------------------------|
| | Qwen3-8B | Qwen3-8B-INT4 |
| Peak Memory (GB) | 16.47 | 6.27 (62% reduction) |
<details>
<summary> Reproduce Peak Memory Usage Results </summary>
We can use the following code to get a sense of peak memory usage during inference:
```Py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig
# use "Qwen/Qwen3-8B" or "pytorch/Qwen3-8B-INT4"
model_id = "pytorch/Qwen3-8B-INT4"
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_id)
torch.cuda.reset_peak_memory_stats()
prompt = "Hey, are you conscious? Can you talk to me?"
messages = [
{
"role": "system",
"content": "",
},
{"role": "user", "content": prompt},
]
templated_prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
print("Prompt:", prompt)
print("Templated prompt:", templated_prompt)
inputs = tokenizer(
templated_prompt,
return_tensors="pt",
).to("cuda")
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128)
output_text = tokenizer.batch_decode(
generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Response:", output_text[0][len(prompt):])
mem = torch.cuda.max_memory_reserved() / 1e9
print(f"Peak Memory Usage: {mem:.02f} GB")
```
</details>
# Model Performance
Our INT4 model is only optimized for batch size 1, so expect some slowdown with larger batch sizes, we expect this to be used in local server deployment for single or a few users where the decode tokens per second will matters more than the time to first token.
## Results (A100 machine)
| Benchmark (Latency) | | |
|----------------------------------|----------------|--------------------------|
| | Qwen3-8B | Qwen3-8B-INT4 |
| latency (batch_size=1) | 3.52s | 2.84s (1.24x speedup) |
Int4 weight only is optimized for batch size 1 and short input and output token length, please stay tuned for models optimized for larger batch sizes or longer token length.
<details>
<summary> Reproduce Model Performance Results </summary>
## Setup
Get vllm source code:
```Shell
git clone git@github.com:vllm-project/vllm.git
```
Install vllm
```
VLLM_USE_PRECOMPILED=1 pip install --editable .
```
Run the benchmarks under `vllm` root folder:
## benchmark_latency
### baseline
```Shell
export MODEL=Qwen/Qwen3-8B
python benchmarks/benchmark_latency.py --input-len 256 --output-len 256 --model $MODEL --batch-size 1
```
### INT4
```Shell
export MODEL=pytorch/Qwen3-8B-INT4
VLLM_DISABLE_COMPILE_CACHE=1 python benchmarks/benchmark_latency.py --input-len 256 --output-len 256 --model $MODEL --batch-size 1
```
## benchmark_serving
We benchmarked the throughput in a serving environment.
Download sharegpt dataset:
```Shell
wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
```
Other datasets can be found in: https://github.com/vllm-project/vllm/tree/main/benchmarks
Note: you can change the number of prompts to be benchmarked with `--num-prompts` argument for `benchmark_serving` script.
### baseline
Server:
```Shell
export MODEL=Qwen/Qwen3-8B
vllm serve $MODEL --tokenizer $MODEL -O3
```
Client:
```Shell
export MODEL=Qwen/Qwen3-8B
python benchmarks/benchmark_serving.py --backend vllm --dataset-name sharegpt --tokenizer $MODEL --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json --model $MODEL --num-prompts 1
```
### INT4
Server:
```Shell
export MODEL=pytorch/Qwen3-8B-INT4
VLLM_DISABLE_COMPILE_CACHE=1 vllm serve $MODEL --tokenizer $MODEL -O3 --pt-load-map-location cuda:0
```
Client:
```Shell
export MODEL=pytorch/Qwen3-8B-INT4
python benchmarks/benchmark_serving.py --backend vllm --dataset-name sharegpt --tokenizer $MODEL --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json --model $MODEL --num-prompts 1
```
</details>
# Paper: TorchAO: PyTorch-Native Training-to-Serving Model Optimization
The model's quantization is powered by **TorchAO**, a framework presented in the paper [TorchAO: PyTorch-Native Training-to-Serving Model Optimization](https://huggingface.co/papers/2507.16099).
**Abstract:** We present TorchAO, a PyTorch-native model optimization framework leveraging quantization and sparsity to provide an end-to-end, training-to-serving workflow for AI models. TorchAO supports a variety of popular model optimization techniques, including FP8 quantized training, quantization-aware training (QAT), post-training quantization (PTQ), and 2:4 sparsity, and leverages a novel tensor subclass abstraction to represent a variety of widely-used, backend agnostic low precision data types, including INT4, INT8, FP8, MXFP4, MXFP6, and MXFP8. TorchAO integrates closely with the broader ecosystem at each step of the model optimization pipeline, from pre-training (TorchTitan) to fine-tuning (TorchTune, Axolotl) to serving (HuggingFace, vLLM, SGLang, ExecuTorch), connecting an otherwise fragmented space in a single, unified workflow. TorchAO has enabled recent launches of the quantized Llama 3.2 1B/3B and LlamaGuard3-8B models and is open-source at this https URL .
# Resources
* **Official TorchAO GitHub Repository:** [https://github.com/pytorch/ao](https://github.com/pytorch/ao)
* **TorchAO Documentation:** [https://docs.pytorch.org/ao/stable/index.html](https://docs.pytorch.org/ao/stable/index.html)
# Disclaimer
PyTorch has not performed safety evaluations or red teamed the quantized models. Performance characteristics, outputs, and behaviors may differ from the original models. Users are solely responsible for selecting appropriate use cases, evaluating and mitigating for accuracy, safety, and fairness, ensuring security, and complying with all applicable laws and regulations.
Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the licenses the models are released under, including any limitations of liability or disclaimers of warranties provided therein.
|
goptouy/blockassist-bc-dappled_leaping_anaconda_1756490231
|
goptouy
| 2025-08-29T17:57:29Z | 0 | 0 | null |
[
"gensyn",
"blockassist",
"gensyn-blockassist",
"minecraft",
"dappled leaping anaconda",
"arxiv:2504.07091",
"region:us"
] | null | 2025-08-29T17:57:12Z |
---
tags:
- gensyn
- blockassist
- gensyn-blockassist
- minecraft
- dappled leaping anaconda
---
# Gensyn BlockAssist
Gensyn's BlockAssist is a distributed extension of the paper [AssistanceZero: Scalably Solving Assistance Games](https://arxiv.org/abs/2504.07091).
|
OpenGVLab/InternVL3_5-1B-Pretrained
|
OpenGVLab
| 2025-08-29T17:57:08Z | 84 | 1 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternViT-300M-448px-V2_5",
"base_model:merge:OpenGVLab/InternViT-300M-448px-V2_5",
"base_model:Qwen/Qwen3-0.6B",
"base_model:merge:Qwen/Qwen3-0.6B",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:49Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternViT-300M-448px-V2_5
- Qwen/Qwen3-0.6B
base_model_relation: merge
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-1B-Pretrained
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
OpenGVLab/InternVL3_5-8B
|
OpenGVLab
| 2025-08-29T17:57:06Z | 2,594 | 32 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternVL3_5-8B-MPO",
"base_model:finetune:OpenGVLab/InternVL3_5-8B-MPO",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:47Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternVL3_5-8B-MPO
base_model_relation: finetune
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-8B
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
OpenGVLab/InternVL3_5-8B-MPO
|
OpenGVLab
| 2025-08-29T17:57:06Z | 68 | 2 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternVL3_5-8B-Instruct",
"base_model:finetune:OpenGVLab/InternVL3_5-8B-Instruct",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:39Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternVL3_5-8B-Instruct
base_model_relation: finetune
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-8B-MPO
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
OpenGVLab/InternVL3_5-8B-Instruct
|
OpenGVLab
| 2025-08-29T17:57:06Z | 319 | 6 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternVL3_5-8B-Pretrained",
"base_model:finetune:OpenGVLab/InternVL3_5-8B-Pretrained",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:39Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternVL3_5-8B-Pretrained
base_model_relation: finetune
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-8B-Instruct
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
OpenGVLab/InternVL3_5-4B
|
OpenGVLab
| 2025-08-29T17:57:05Z | 1,253 | 18 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternVL3_5-4B-MPO",
"base_model:finetune:OpenGVLab/InternVL3_5-4B-MPO",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:39Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternVL3_5-4B-MPO
base_model_relation: finetune
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-4B
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
OpenGVLab/InternVL3_5-241B-A28B-MPO
|
OpenGVLab
| 2025-08-29T17:57:02Z | 178 | 2 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternVL3_5-241B-A28B-Instruct",
"base_model:finetune:OpenGVLab/InternVL3_5-241B-A28B-Instruct",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:34Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternVL3_5-241B-A28B-Instruct
base_model_relation: finetune
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-241B-A28B-MPO
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
OpenGVLab/InternVL3_5-30B-A3B-Pretrained
|
OpenGVLab
| 2025-08-29T17:57:02Z | 58 | 1 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternViT-300M-448px-V2_5",
"base_model:merge:OpenGVLab/InternViT-300M-448px-V2_5",
"base_model:Qwen/Qwen3-30B-A3B",
"base_model:merge:Qwen/Qwen3-30B-A3B",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:37Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternViT-300M-448px-V2_5
- Qwen/Qwen3-30B-A3B
base_model_relation: merge
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-30B-A3B-Pretrained
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
OpenGVLab/InternVL3_5-38B
|
OpenGVLab
| 2025-08-29T17:57:02Z | 1,099 | 22 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternVL3_5-38B-MPO",
"base_model:finetune:OpenGVLab/InternVL3_5-38B-MPO",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:37Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternVL3_5-38B-MPO
base_model_relation: finetune
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-38B
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
OpenGVLab/InternVL3_5-241B-A28B-Instruct
|
OpenGVLab
| 2025-08-29T17:57:02Z | 43 | 13 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternVL3_5-241B-A28B-Pretrained",
"base_model:finetune:OpenGVLab/InternVL3_5-241B-A28B-Pretrained",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:34Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternVL3_5-241B-A28B-Pretrained
base_model_relation: finetune
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-241B-A28B-Instruct
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
OpenGVLab/InternVL3_5-30B-A3B-MPO
|
OpenGVLab
| 2025-08-29T17:57:02Z | 69 | 3 |
transformers
|
[
"transformers",
"safetensors",
"internvl_chat",
"feature-extraction",
"internvl",
"custom_code",
"image-text-to-text",
"conversational",
"multilingual",
"dataset:OpenGVLab/MMPR-v1.2",
"dataset:OpenGVLab/MMPR-Tiny",
"arxiv:2312.14238",
"arxiv:2404.16821",
"arxiv:2412.05271",
"arxiv:2411.10442",
"arxiv:2504.10479",
"arxiv:2508.18265",
"base_model:OpenGVLab/InternVL3_5-30B-A3B-Instruct",
"base_model:finetune:OpenGVLab/InternVL3_5-30B-A3B-Instruct",
"license:apache-2.0",
"region:us"
] |
image-text-to-text
| 2025-08-25T16:38:37Z |
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
- OpenGVLab/InternVL3_5-30B-A3B-Instruct
base_model_relation: finetune
datasets:
- OpenGVLab/MMPR-v1.2
- OpenGVLab/MMPR-Tiny
language:
- multilingual
tags:
- internvl
- custom_code
---
# InternVL3_5-30B-A3B-MPO
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) [\[📜 InternVL2.5-MPO\]](https://huggingface.co/papers/2411.10442) [\[📜 InternVL3\]](https://huggingface.co/papers/2504.10479) [\[📜 InternVL3.5\]](https://huggingface.co/papers/2508.18265)
[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://chat.intern-ai.org.cn/) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
<div align="center">
<img width="500" alt="image" src="https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/zJsd2hqd3EevgXo6fNgC-.png">
</div>
## Introduction
We introduce *InternVL3.5*, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the *Cascade Reinforcement Learning (Cascade RL)* framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a *Visual Resolution Router (ViR)* that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled *Vision-Language Deployment (DvD)* strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05 \\(\times\\) inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks—narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
> Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and agentic benchmarks: MMBench v1.1 (en), MMStar,BLINK, HallusionBench, AI2D, OCRBench, MMVet, MME-RealWorld (en), MVBench, VideoMME, MMMU, MathVista, MathVision, MathVerse, DynaMath, WeMath, LogicVista, MATH500, AIME24, AIME25, GPQA, MMLU-Pro, GAOKAO, IFEval, SGP-Bench, VSI-Bench, ERQA, SpaCE-10, and OmniSpatial.
See [quick start](#quick-start) for how to use our model.
## InternVL3.5 Family
In the following table, we provide an overview of the InternVL3.5 series.
To maintain consistency with earlier generations, we provide two model formats: [the GitHub format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B), consistent with prior releases, and [the HF format](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF), aligned with the official Transformers standard.
> If you want to convert the checkpoint between these two formats, please refer to the scripts about [custom2hf](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_custom2hf.py) and [hf2custom](https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/tools/internvl_hf2custom.py).
### Github Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| --------------------- | ------------- | --------------- | ------------ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| InternVL3.5-1B | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-38B | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-20B-A4B | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview) |
| InternVL3.5-30B-A3B | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-241B-A28B | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
### HuggingFace Format
| Model | #Vision Param | #Language Param | #Total Param | HF Link | ModelScope Link |
| ------------------------ | ------------- | --------------- | ------------ | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| InternVL3.5-1B-HF | 0.3B | 0.8B | 1.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-HF) |
| InternVL3.5-2B-HF | 0.3B | 2.0B | 2.3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-HF) |
| InternVL3.5-4B-HF | 0.3B | 4.4B | 4.7B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-HF) |
| InternVL3.5-8B-HF | 0.3B | 8.2B | 8.5B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-HF) |
| InternVL3.5-14B-HF | 0.3B | 14.8B | 15.1B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-HF) |
| InternVL3.5-38B-HF | 5.5B | 32.8B | 38.4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-HF) |
| InternVL3.5-20B-A4B-HF | 0.3B | 20.9B | 21.2B-A4B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |
| InternVL3.5-30B-A3B-HF | 0.3B | 30.5B | 30.8B-A3B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-HF) |
| InternVL3.5-241B-A28B-HF | 5.5B | 235.1B | 240.7B-A28B | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-HF) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-HF) |
> We conduct the evaluation with [VLMEvalkit](https://github.com/open-compass/VLMEvalKit). ***To enable the Thinking mode of our model, please set the system prompt to [R1_SYSTEM_PROMPT](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/vlm/internvl/internvl_chat.py#L38).*** When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
Our training pipeline comprises four stages: Multimodal Continual Pre-Training (**CPT**), Supervised Fine-Tuning (**SFT**), and Cascade Reinforcement Learning (**CascadeRL**). In CascadeRL, we first fine-tune the model using Mixed Preference Optimization (**MPO**) under an offline RL setting, followed by **GSPO** under an oneline RL setting.
For the Flash version of InternVL3.5, we additionally introduce a lightweight training stage, termed Visual Consistency Learning (**ViCO**), which reduces the token cost required to represent an image patch.
Here, we also open-source the model weights after different training stages for potential research usage.
***If you're unsure which version to use, please select the one without any suffix, as it has completed the full training pipeline.***
| Model | Training Pipeline | HF Link | ModelScope Link |
| -------------------------------- | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- |
| InternVL3.5-1B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Pretrained) |
| InternVL3.5-1B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-Instruct) |
| InternVL3.5-1B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B-MPO) |
| InternVL3.5-1B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-1B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-1B) |
| InternVL3.5-2B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Pretrained) |
| InternVL3.5-2B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-Instruct) |
| InternVL3.5-2B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B-MPO) |
| InternVL3.5-2B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-2B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-2B) |
| InternVL3.5-4B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Pretrained) |
| InternVL3.5-4B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-Instruct) |
| InternVL3.5-4B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B-MPO) |
| InternVL3.5-4B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-4B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-4B) |
| InternVL3.5-8B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Pretrained) |
| InternVL3.5-8B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-Instruct) |
| InternVL3.5-8B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B-MPO) |
| InternVL3.5-8B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-8B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-8B) |
| InternVL3.5-14B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Pretrained) |
| InternVL3.5-14B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-Instruct) |
| InternVL3.5-14B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B-MPO) |
| InternVL3.5-14B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-14B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-14B) |
| InternVL3.5-30B-A3B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Pretrained) |
| InternVL3.5-30B-A3B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-Instruct) |
| InternVL3.5-30B-A3B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B-MPO) |
| InternVL3.5-30B-A3B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-30B-A3B) |
| InternVL3.5-38B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Pretrained) |
| InternVL3.5-38B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-Instruct) |
| InternVL3.5-38B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B-MPO) |
| InternVL3.5-38B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-38B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-38B) |
| InternVL3.5-241B-A28B-Pretrained | CPT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Pretrained) |
| InternVL3.5-241B-A28B-Instruct | CPT + SFT | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-Instruct) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-Instruct) |
| InternVL3.5-241B-A28B-MPO | CPT + SFT + MPO | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B-MPO) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B-MPO) |
| InternVL3.5-241B-A28B | CPT + SFT + CascadeRL | [🤗 link](https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B) | [🤖 link](https://www.modelscope.cn/models/OpenGVLab/InternVL3_5-241B-A28B) |
The Flash version of our model will be released as soon as possible.
## Model Architecture
`InternVL3.5`:
This series of models follow the "ViT–MLP–LLM" paradigm adopted in previous versions of InternVL.
We initialize the language model using the Qwen3 series and GPT-OSS, and the vision encoder using InternViT-300M and InternViT-6B.
The Dynamic High Resolution strategy introduced in InternVL1.5 is also retained in our design.
`InternVL3.5-Flash`:
Compared to InternVL3.5, InternVL3.5-Flash further integrates the *Visual Resolution Router (ViR)*, thus yielding a series of efficient variants friendly suitable for resource-constrained scenarios.
Specifically, in InternVL3.5, each image patch is initially represented as 1024 visual tokens for the vision encoder, which are then compressed into 256 tokens via a pixel shuffle module before being passed to the Large Language Model (LLM).
In InternVL3.5-Flash, as shown in the Figure below, an additional pixel shuffle module with a higher compression rate is included, enabling the compression of visual tokens down to 64 tokens.
For each patch, the patch router determines the appropriate compression rate by assessing its semantic richness, and routes it to the corresponding pixel shuffle module accordingly.
Benefiting from this patch-aware compression mechanism, InternVL3.5-Flash is able to reduce the number of visual tokens by 50\% while maintaining nearly 100\% of the performance of InternVL3.5.
## Training and Deployment Strategy
### Pre-Training
During the pre-training stage, we update all model parameters jointly using the combination of large-scale text and multimodal corpora. Specifically, given an arbitrary training sample consisting of a multimodal token sequence \\(\mathbf{x}=\left(x_1, x_2, \ldots, x_L\right)\\), the next token prediction (NTP) loss is calculated on each text token as follows:
$$
\mathcal{L}_{i}=-\log p_\theta\left(x_i \mid x_1, \ldots, x_{i-1}\right),
$$
where \\(x_i\\) is the predicted token and prefix tokens in \\(\{x_1, x_2, \ldots, x_{i-1}\}\\) can be either text tokens or image tokens. Notably, for conversation samples, only response tokens are included for the calculation of the loss.
Additionally, to mitigate bias toward either longer or shorter responses during training, we adopt the square averaging to re-weight the NTP loss as follows:
$$
\mathcal{L}_{i}^{'} = \frac{w_i}{\sum_j w_j} \cdot \mathcal{L}_i, \quad w_i = \frac{1}{N^{0.5}},
$$
where \\(N\\) denotes the number of tokens in the training sample on which the loss needs to be calculated. The random JPEG compression is also included to enhance the model's real-world performance.
### Supervised Fine-Tuning
During the SFT phase, we adopt the same objective as in the pre-training stage and use the square-root averaging strategy to calculate the final loss. In this stage, the context window is set to 32K tokens to adapt long-context information.
Compared to InternVL3, the SFT stage of InternVL3.5 contains more high-quality and diverse training data derived from three sources:
(1) Instruction-following data from InternVL3, which are reused to preserve broad coverage of vision–language tasks.
(2) Multimodal reasoning data in the "Thinking" mode, which are included to instill long-thinking capabilities in the model. To construct such data, we first use InternVL3-78B to describe the image and then input the description into DeepSeek-R1 to sample rollouts with detailed reasoning processes. Rollouts with an incorrect final answer are filtered out. The questions in these datasets cover various expert domains, such as mathematics and scientific disciplines, thereby strengthening performance on different reasoning tasks.
(3) Capability-expansion datasets, which endow InternVL3.5 with new skills, including GUI-based interaction, embodied interaction, and scalable vect
### Cascade Reinforcement Learning
Cascade RL aims to combine the benefits of offline RL and online RL to progressively facilitate the post-training of MLLMs in an efficient manner.
Specifically, we first fine-tune the model using an offline RL algorithm as an efficient warm-up stage to reach a satisfied results, which can guarantee the high-quality rollouts for the latter stage.
Subsequently, we employ an online RL algorithm to further refine the output distribution based on rollouts generated by the model itself. Compared to the single offline or online RL stage, our cascaded RL achieves significant performance improvements at a fraction of the GPU time cost.
During the offline RL stage, we employ mixed preference optimization (MPO) to fine-tune the model. Specifically, the training objective of MPO is a combination of preference loss \\(\mathcal{L}_{p}\\), quality loss \\(\mathcal{L}_{q}\\), and generation loss \\(\mathcal{L}_{g}\\), which can be formulated as follows:
$$
\mathcal{L}_{\text{MPO}}=
w_{p} \mathcal{L}_{p}
+
w_{q} \mathcal{L}_{q}
+
w_{g} \mathcal{L}_{g}
,
$$
where \\(w_{*}\\) represents the weight assigned to each loss component.
The DPO loss, BCO loss, and LM loss serve as the preference loss, quality loss, and generation loss, respectively.
During the online RL stage, we employ GSPO, without reference model constraints, as our online RL algorithm, which we find more effective in training both dense and mixture-of-experts (MoE) models. Similar to GRPO, the advantage is defined as the normalized reward across responses sampled from the same query.
The training objective of GSPO is given by:
$$
\mathcal{L}_{\mathrm{GSPO}}(\theta)=\mathbb{E}_{x \sim \mathcal{D},\left\{y_i\right\}_{i=1}^G \sim \pi_{\theta \text { old }}(\cdot \mid x)}\left[\frac{1}{G} \sum_{i=1}^G \min \left(s_i(\theta) \widehat{A}_i, \operatorname{clip}\left(s_i(\theta), 1-\varepsilon, 1+\varepsilon\right) \widehat{A}_i\right)\right],
$$
where the importance sampling ratio is defined as the geometric mean of the per-token ratios.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Visual Consistency Learning
We further include ViCO as an additional training stage to integrate the *visual resolution router (ViR)* into InternVL3.5, thereby reducing the inference cost of InternVL3.5. The obtained efficient version of InternVL3.5 are termed as *InternVL3.5-Flash*. In particular, ViCO comprises two stages:
`Consistency training`:
In this stage, the entire model is trained to minimize the divergence between response distributions conditioned on visual tokens with different compression rates.
In practice, we introduce an extra reference model, which is frozen and initialized with InternVL3.5.
Given a sample, each image patch is represented as either 256 or 64 tokens, and the training objective is defined as follows:
$$
\mathcal{L}_\text{ViCO} =
\mathbb{E}_{\xi \sim \mathcal{R}} \Bigg[
\frac{1}{N} \sum_{i=1}^{N} \mathrm{KL} \Big(
\pi_{\theta_{ref}}\left(y_i \mid y_{<i}, I\right) \;\Big\|\;
\pi_{\theta_{policy}}\left(y_i \mid y_{<i}, I_\xi\right)
\Big)
\Bigg],
$$
where \\(\mathrm{KL}\) denotes the KL divergence and \(\xi\) denotes the compression rate, which is uniformly sampled from \(\{\frac{1}{4},\frac{1}{16}\}\). The image \(I_\xi\) is represented as 256 tokens when \(\xi=\frac{1}{4}\) and 64 tokens when \(\xi=\frac{1}{16}\). Notably, the reference model always performs inference with \(\xi=\frac{1}{4}\).
`Router training`:
This stage aims to train the ViR to select an appropriate trade-off resolution for different inputs.
ViR is formulated as a binary classifier and trained using standard cross-entropy loss.
To construct the route targets, we first compute the KL divergence between the model outputs conditioned on uncompressed visual tokens (i.e., 256 tokens per patch) and those conditioned on compressed visual tokens (i.e., 64 tokens per patch).
During this stage, the main MLLM (ViT, MLP and LLM) is kept frozen, and only the ViR is trained.
Specifically, we first compute the loss ratio for each patch:
$$
r_i = \frac{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{16}}\big)}{\mathcal{L}_\text{ViCO}\big(y_i \mid I_{\frac{1}{4}}\big)},
$$
which quantifies the relative increase in loss caused by compressing the visual tokens. Based on this ratio, the binary ground-truth label for the patch router is defined as:
$$
y_i^\text{router} =
\begin{cases}
0, & r_i < \tau \; \text{(compression has negligible impact)} \\
1, & r_i \ge \tau \; \text{(compression has significant impact)},
\end{cases}
$$
where \(y_i^{\text{router}}=0\) and \(y_i^{\text{router}}=1\) indicate that the compression rate \(\xi\) is set to \(\tfrac{1}{16}\) and \(\tfrac{1}{4}\), respectively.
> Please see [our paper](https://huggingface.co/papers/2508.18265) for more technical and experimental details.
### Test-Time Scaling
Test-time scaling (TTS) has been empirically demonstrated as an effective approach to enhance the reasoning capabilities of LLMs and MLLMs, particularly for complex tasks necessitating multi-step inference.
In this work, we implement a comprehensive test-time scaling approach that simultaneously improves reasoning depth (i.e., deep thinking) and breadth (i.e., parallel thinking).
`Deep Thinking`: By activating the Thinking mode, we guide the model to deliberately engage in step-by-step reasoning (i.e., decomposing complex problems into logical steps and validating intermediate conclusions) prior to generating the final answer. This approach systematically improves the logical structure of solutions for complex problems, particularly those requiring multi-step inference, and enhances reasoning depth.
`Parallel Thinking`: Following InternVL3, for reasoning tasks, we adopt the Best-of-N (BoN) strategy by employing [VisualPRM-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1_1) as the critic model to select the optimal response from multiple reasoning candidates.
This approach improves reasoning breadth.
> Notably, unless otherwise specified, the experimental results reported in our paper are obtained without applying TTS. Thus far, we have only applied TTS to reasoning benchmarks, since we found that the model already exhibits strong perception and understanding capabilities, and initiating TTS yields no significant improvement.
### Decoupled Vision-Language Deployment
In multimodal inference, the vision encoder and language model have distinct computational characteristics. The vision encoder that transforms images into semantic features is highly parallelizable and does not rely on long-term history state. In contrast, the language model adopts the inference in an autoregressive manner, which requires previous states to compute the next one. This sequential property makes the language part more sensitive to memory bandwidth and latency.
When MLLMs are deployed online at scale, the vision and language models often block each other, thus incurring additional inference cost. This effect becomes more pronounced with larger vision models or higher-resolution images.
As shown in the Figure above, we propose decoupled vision-language deployment (DvD) to address this issue by separating vision and language processing, with a particular focus on optimizing the prefilling stage. The vision subsystem batches and processes images to produce compact feature embeddings, which are then transmitted to the language subsystem for fusion with the text context prior to decoding. This separation alleviates blocking and brings multimodal prefilling performance closer to that of pure language models.
In our system implementation, the ViT and MLP (and ViR for InternVL3.5-Flash) are deployed on the vision server, while the language server executes only the LLM. The communication is unidirectional, transmitting BF16 visual features over TCP, with RDMA optionally employed to achieve higher transmission speed. Vision processing, feature transmission, and language processing are organized into an asynchronous three-stage pipeline, enabling overlapped execution and minimizing pipeline stalls.
DvD increases GPU utilization and processing efficiency on the vision side, while enabling the language server to focus exclusively on the LLM’s prefilling and decoding without being blocked by vision computation. This design leads to improved throughput and responsiveness. Moreover, the architecture supports independent hardware cost optimization for the vision and language modules, and facilitates the seamless integration of new modules without requiring modifications to the language server deployment.
## Evaluation on Multimodal Capability
### Multimodal Reasoning and Mathematics
### OCR, Chart, and Document Understanding
### Multi-Image Understanding & Real-World Comprehension
### Comprehensive Multimodal Understanding & Multimodal Hallucination Evaluation
### Visual Grounding
### Multimodal Multilingual Understanding
### Video Understanding
### GUI Tasks
### Embodied Tasks
### SVG Tasks
## Evaluation on Language Capability
## Ablation Study
### Cascade Reinforcement Learning
### Decoupled Vision-Language Deployment
## Quick Start
We provide an example code to run `InternVL3.5-8B` using `transformers`. Please note that our models with up to 30B parameters can be deployed on a single A100 GPU, while the 38B model requires two A100 GPUs and the 235B model requires eight A100 GPUs.
> In most cases, both [LMDeploy](https://github.com/InternLM/lmdeploy) and [vLLM](https://github.com/vllm-project/vllm) can be used for model deployment. However, for InternVL3.5-20B-A4B, we recommend using vLLM since lmdeploy has not yet supported GPT-OSS.
> Please use transformers>=4.52.1 to ensure the model works normally. For the 20B version of our model, transformers>=4.55.0 is required.
### Model Loading
#### 16-bit (bf16 / fp16)
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval().cuda()
```
#### BNB 8-bit Quantization
```python
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=True,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True).eval()
```
#### Multiple GPUs
```python
import math
import torch
from transformers import AutoTokenizer, AutoModel
path = "OpenGVLab/InternVL3_5-8B"
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
```
### Thinking Mode
To enable thinking mode, please set the system prompt to our Thinking System Prompt. When enabling Thinking mode, we recommend setting `do_sample=True` and `temperature=0.6` to mitigate undesired repetition.
```python
R1_SYSTEM_PROMPT = """
You are an AI assistant that rigorously follows this response protocol:
1. First, conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step. Enclose this entire thinking process within <think> and </think> tags.
2. After the thinking section, provide a clear, concise, and direct answer to the user's question. Separate the answer from the think section with a newline.
Ensure that the thinking process is thorough but remains focused on the query. The final answer should be standalone and not reference the thinking section.
""".strip()
model.system_message = R1_SYSTEMP_PROMPT
```
### Inference with Transformers
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
def build_transform(input_size):
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
T.ToTensor(),
T.Normalize(mean=MEAN, std=STD)
])
return transform
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
best_ratio_diff = float('inf')
best_ratio = (1, 1)
area = width * height
for ratio in target_ratios:
target_aspect_ratio = ratio[0] / ratio[1]
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
if ratio_diff < best_ratio_diff:
best_ratio_diff = ratio_diff
best_ratio = ratio
elif ratio_diff == best_ratio_diff:
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
best_ratio = ratio
return best_ratio
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
orig_width, orig_height = image.size
aspect_ratio = orig_width / orig_height
# calculate the existing image aspect ratio
target_ratios = set(
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
i * j <= max_num and i * j >= min_num)
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
# find the closest aspect ratio to the target
target_aspect_ratio = find_closest_aspect_ratio(
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
# calculate the target width and height
target_width = image_size * target_aspect_ratio[0]
target_height = image_size * target_aspect_ratio[1]
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
# resize the image
resized_img = image.resize((target_width, target_height))
processed_images = []
for i in range(blocks):
box = (
(i % (target_width // image_size)) * image_size,
(i // (target_width // image_size)) * image_size,
((i % (target_width // image_size)) + 1) * image_size,
((i // (target_width // image_size)) + 1) * image_size
)
# split the image
split_img = resized_img.crop(box)
processed_images.append(split_img)
assert len(processed_images) == blocks
if use_thumbnail and len(processed_images) != 1:
thumbnail_img = image.resize((image_size, image_size))
processed_images.append(thumbnail_img)
return processed_images
def load_image(image_file, input_size=448, max_num=12):
image = Image.open(image_file).convert('RGB')
transform = build_transform(input_size=input_size)
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(image) for image in images]
pixel_values = torch.stack(pixel_values)
return pixel_values
path = 'OpenGVLab/InternVL3_5-8B'
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
load_in_8bit=False,
low_cpu_mem_usage=True,
use_flash_attn=True,
trust_remote_code=True,
device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)
# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list,
history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
num_patches_list=num_patches_list,
questions=questions,
generation_config=generation_config)
for question, response in zip(questions, responses):
print(f'User: {question}\nAssistant: {response}')
# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
if bound:
start, end = bound[0], bound[1]
else:
start, end = -100000, 100000
start_idx = max(first_idx, round(start * fps))
end_idx = min(round(end * fps), max_frame)
seg_size = float(end_idx - start_idx) / num_segments
frame_indices = np.array([
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
for idx in range(num_segments)
])
return frame_indices
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
max_frame = len(vr) - 1
fps = float(vr.get_avg_fps())
pixel_values_list, num_patches_list = [], []
transform = build_transform(input_size=input_size)
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
for frame_index in frame_indices:
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
pixel_values = [transform(tile) for tile in img]
pixel_values = torch.stack(pixel_values)
num_patches_list.append(pixel_values.shape[0])
pixel_values_list.append(pixel_values)
pixel_values = torch.cat(pixel_values_list)
return pixel_values, num_patches_list
video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')
question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')
```
#### Streaming Output
Besides this method, you can also use the following code to get streamed output.
```python
from transformers import TextIteratorStreamer
from threading import Thread
# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
history=None, return_history=False, generation_config=generation_config,
))
thread.start()
# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
if new_text == model.conv_template.sep:
break
generated_text += new_text
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
```
## Finetune
Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTuner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
## Deployment
### LMDeploy
LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs.
```sh
pip install lmdeploy>=0.9.1
```
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
#### A 'Hello, world' Example
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
response = pipe(('describe this image', image))
print(response.text)
```
#### Multi-images Inference
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
from lmdeploy.vl.constants import IMAGE_TOKEN
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
]
images = [load_image(img_url) for img_url in image_urls]
# Numbering images improves multi-image conversations
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
print(response.text)
```
#### Batch Prompts Inference
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
```python
from lmdeploy import pipeline, PytorchEngineConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image_urls=[
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
]
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
response = pipe(prompts)
print(response)
```
#### Multi-turn Conversation
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
```python
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig
from lmdeploy.vl import load_image
# Please set tp=2 for the 38B version and tp=8 for the 241B-A28B version.
model = 'OpenGVLab/InternVL3_5-8B'
pipe = pipeline(model, backend_config=PytorchEngineConfig(session_len=32768, tp=1))
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
gen_config = GenerationConfig(top_k=50, top_p=0.95, temperature=0.6, max_new_tokens=8192)
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
print(sess.response.text)
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
print(sess.response.text)
```
#### Service
LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
```shell
lmdeploy serve api_server OpenGVLab/InternVL3_5-8B --server-port 23333 --tp 1 --backend pytorch
```
To use the OpenAI-style interface, you need to install OpenAI:
```shell
pip install openai
```
Then, use the code below to make the API call:
```python
from openai import OpenAI
client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
model=model_name,
messages=[{
'role':
'user',
'content': [{
'type': 'text',
'text': 'describe this image',
}, {
'type': 'image_url',
'image_url': {
'url':
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
},
}],
}],
temperature=0.8,
top_p=0.8)
print(response)
```
## License
This project is released under the apache-2.0 License. This project uses the pre-trained Qwen3 as a component, which is licensed under the apache-2.0 License.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@article{wang2025internvl3_5,
title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
journal={arXiv preprint arXiv:2508.18265},
year={2025}
}
```
|
Subsets and Splits
Filtered Qwen2.5 Distill Models
Identifies specific configurations of models by filtering cards that contain 'distill', 'qwen2.5', '7b' while excluding certain base models and incorrect model ID patterns, uncovering unique model variants.
Filtered Model Cards Count
Finds the count of entries with specific card details that include 'distill', 'qwen2.5', '7b' but exclude certain base models, revealing valuable insights about the dataset's content distribution.
Filtered Distill Qwen 7B Models
Filters for specific card entries containing 'distill', 'qwen', and '7b', excluding certain strings and patterns, to identify relevant model configurations.
Filtered Qwen-7b Model Cards
The query performs a detailed filtering based on specific keywords and excludes certain entries, which could be useful for identifying a specific subset of cards but does not provide deeper insights or trends.
Filtered Qwen 7B Model Cards
The query filters for specific terms related to "distilled" or "distill", "qwen", and "7b" in the 'card' column but excludes certain base models, providing a limited set of entries for further inspection.
Qwen 7B Distilled Models
The query provides a basic filtering of records to find specific card names that include keywords related to distilled Qwen 7b models, excluding a particular base model, which gives limited insight but helps in focusing on relevant entries.
Qwen 7B Distilled Model Cards
The query filters data based on specific keywords in the modelId and card fields, providing limited insight primarily useful for locating specific entries rather than revealing broad patterns or trends.
Qwen 7B Distilled Models
Finds all entries containing the terms 'distilled', 'qwen', and '7b' in a case-insensitive manner, providing a filtered set of records but without deeper analysis.
Distilled Qwen 7B Models
The query filters for specific model IDs containing 'distilled', 'qwen', and '7b', providing a basic retrieval of relevant entries but without deeper analysis or insight.
Filtered Model Cards with Distill Qwen2.
Filters and retrieves records containing specific keywords in the card description while excluding certain phrases, providing a basic count of relevant entries.
Filtered Model Cards with Distill Qwen 7
The query filters specific variations of card descriptions containing 'distill', 'qwen', and '7b' while excluding a particular base model, providing limited but specific data retrieval.
Distill Qwen 7B Model Cards
The query filters and retrieves rows where the 'card' column contains specific keywords ('distill', 'qwen', and '7b'), providing a basic filter result that can help in identifying specific entries.