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">
       &nbsp&nbsp🤗 <a href="https://huggingface.co/PowerInfer">Hugging Face</a>&nbsp&nbsp | &nbsp&nbsp🤖 <a href="https://modelscope.cn/organization/PowerInfer">ModelScope</a>&nbsp&nbsp | &nbsp&nbsp 📑 <a href="https://github.com/SJTU-IPADS/SmallThinker/blob/main/smallthinker-technical-report.pdf">Technical Report</a> &nbsp&nbsp
       &nbsp&nbsp 📚 <a href="https://huggingface.co/papers/2507.20984">Paper</a> &nbsp&nbsp | &nbsp&nbsp 💻 <a href="https://github.com/SJTU-IPADS/SmallThinker">GitHub Repo</a> &nbsp&nbsp
</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.