File size: 6,988 Bytes
003b904 4859198 9df52ed 0ad4904 4859198 0ad4904 003b904 b5bd753 003b904 b5bd753 8c5b3a3 0ad4904 8c5b3a3 b5bd753 003b904 e726dc8 003b904 9df52ed 003b904 7883805 0ad4904 003b904 109a2c4 eb58b5c 003b904 d2557f8 0ad4904 d2557f8 fe1da60 b95a223 3abcff8 003b904 eddd5e2 003b904 9b05bde |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 |
---
language:
- en
library_name: transformers
license: apache-2.0
pipeline_tag: text-generation
tags:
- moe
---
## Introduction
<p align="center">
  🤗 <a href="https://huggingface.co/PowerInfer">Hugging Face</a>   |   🤖 <a href="https://modelscope.cn/organization/PowerInfer">ModelScope</a>   |    📑 <a href="https://github.com/SJTU-IPADS/SmallThinker/blob/main/smallthinker-technical-report.pdf">Technical Report</a>   
   📚 <a href="https://huggingface.co/papers/2507.20984">Paper</a>    |    💻 <a href="https://github.com/SJTU-IPADS/SmallThinker">GitHub Repo</a>   
</p>
SmallThinker is a family of **on-device native** Mixture-of-Experts (MoE) language models specially designed for local deployment,
co-developed by the **IPADS** and **School of AI at Shanghai Jiao Tong University** and **Zenergize AI**.
Designed from the ground up for resource-constrained environments,
SmallThinker brings powerful, private, and low-latency AI directly to your personal devices,
without relying on the cloud.
## Paper
The model was presented in the paper [SmallThinker: A Family of Efficient Large Language Models Natively Trained for Local Deployment](https://huggingface.co/papers/2507.20984).
### Abstract
While frontier large language models (LLMs) continue to push capability boundaries, their deployment remains confined to GPU-powered cloud infrastructure. We challenge this paradigm with SmallThinker, a family of LLMs natively designed - not adapted - for the unique constraints of local devices: weak computational power, limited memory, and slow storage. Unlike traditional approaches that mainly compress existing models built for clouds, we architect SmallThinker from the ground up to thrive within these limitations. Our innovation lies in a deployment-aware architecture that transforms constraints into design principles. First, We introduce a two-level sparse structure combining fine-grained Mixture-of-Experts (MoE) with sparse feed-forward networks, drastically reducing computational demands without sacrificing model capacity. Second, to conquer the I/O bottleneck of slow storage, we design a pre-attention router that enables our co-designed inference engine to prefetch expert parameters from storage while computing attention, effectively hiding storage latency that would otherwise cripple on-device inference. Third, for memory efficiency, we utilize NoPE-RoPE hybrid sparse attention mechanism to slash KV cache requirements. We release SmallThinker-4B-A0.6B and SmallThinker-21B-A3B, which achieve state-of-the-art performance scores and even outperform larger LLMs. Remarkably, our co-designed system mostly eliminates the need for expensive GPU hardware: with Q4_0 quantization, both models exceed 20 tokens/s on ordinary consumer CPUs, while consuming only 1GB and 8GB of memory respectively. SmallThinker is publicly available at this http URL and this http URL .
## Performance
Note: The model is trained mainly on English.
| Model | MMLU | GPQA-diamond | MATH-500 | IFEVAL | LIVEBENCH | HUMANEVAL | Average |
|---|---|---|---|---|---|---|---|
| **SmallThinker-21BA3B-Instruct** | 84.43 | <u>55.05</u> | 82.4 | **85.77** | **60.3** | <u>89.63</u> | **76.26** |
| Gemma3-12b-it | 78.52 | 34.85 | 82.4 | 74.68 | 44.5 | 82.93 | 66.31 |
| Qwen3-14B | <u>84.82</u> | 50 | **84.6** | <u>85.21</u>| <u>59.5</u> | 88.41 | <u>75.42</u> |
| Qwen3-30BA3B | **85.1** | 44.4 | <u>84.4</u> | 84.29 | 58.8 | **90.24** | 74.54 |
| Qwen3-8B | 81.79 | 38.89 | 81.6 | 83.92 | 49.5 | 85.9 | 70.26 |
| Phi-4-14B | 84.58 | **55.45** | 80.2 | 63.22 | 42.4 | 87.2 | 68.84 |
For the MMLU evaluation, we use a 0-shot CoT setting.
All models are evaluated in non-thinking mode.
## Speed
| Model | Memory(GiB) | i9 14900 | 1+13 8ge4 | rk3588 (16G) | Raspberry PI 5 |
|---|---|---|---|---|---|
| SmallThinker 21B+sparse | 11.47 | 30.19 | 23.03 | 10.84 | 6.61 |
| SmallThinker 21B+sparse+limited memory | limit 8G | 20.30 | 15.50 | 8.56 | - |
| Qwen3 30B A3B | 16.20 | 33.52 | 20.18 | 9.07 | - |
| Qwen3 30B A3B+limited memory | limit 8G | 10.11 | 0.18 | 6.32 | - |
| Gemma 3n E2B | 1G, theoretically | 36.88 | 27.06 | 12.50 | 6.66 |
| Gemma 3n E4B | 2G, theoretically | 21.93 | 16.58 | 7.37 | 4.01 |
Note: i9 14900, 1+13 8ge4 use 4 threads, others use the number of threads that can achieve the maximum speed. All models here have been quantized to q4_0.
You can deploy SmallThinker with offloading support using [PowerInfer](https://github.com/SJTU-IPADS/PowerInfer/tree/main/smallthinker)
## Model Card
<div align="center">
| **Architecture** | Mixture-of-Experts (MoE) |
|:---:|:---:|
| **Total Parameters** | 21B |
| **Activated Parameters** | 3B |
| **Number of Layers** | 52 |
| **Attention Hidden Dimension** | 2560 |
| **MoE Hidden Dimension** (per Expert) | 768 |
| **Number of Attention Heads** | 28 |
| **Number of KV Heads** | 4 |
| **Number of Experts** | 64 |
| **Selected Experts per Token** | 6 |
| **Vocabulary Size** | 151,936 |
| **Context Length** | 16K |
| **Attention Mechanism** | GQA |
| **Activation Function** | ReGLU |
</div>
## How to Run
### Transformers
`transformers==4.53.3` is required, we are actively working to support the latest version.
The following contains a code snippet illustrating how to use the model generate content based on given inputs.
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
path = "PowerInfer/SmallThinker-21BA3B-Instruct"
device = "cuda"
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map=device, trust_remote_code=True)
messages = [
{"role": "user", "content": "Give me a short introduction to large language model."},
]
model_inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(device)
model_outputs = model.generate(
model_inputs,
do_sample=True,
max_new_tokens=1024
)
output_token_ids = [
model_outputs[i][len(model_inputs[i]):] for i in range(len(model_inputs))
]
responses = tokenizer.batch_decode(output_token_ids, skip_special_tokens=True)[0]
print(responses)
```
### ModelScope
`ModelScope` adopts Python API similar to (though not entirely identical to) `Transformers`. For basic usage, simply modify the first line of the above code as follows:
```python
from modelscope import AutoModelForCausalLM, AutoTokenizer
```
## Statement
- Due to the constraints of its model size and the limitations of its training data, its responses may contain factual inaccuracies, biases, or outdated information.
- Users bear full responsibility for independently evaluating and verifying the accuracy and appropriateness of all generated content.
- SmallThinker does not possess genuine comprehension or consciousness and cannot express personal opinions or value judgments. |