Instructions to use Infinigence/Megrez2-3x7B-A3B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Infinigence/Megrez2-3x7B-A3B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Infinigence/Megrez2-3x7B-A3B", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Infinigence/Megrez2-3x7B-A3B", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Infinigence/Megrez2-3x7B-A3B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Infinigence/Megrez2-3x7B-A3B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Infinigence/Megrez2-3x7B-A3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Infinigence/Megrez2-3x7B-A3B

SGLang

How to use Infinigence/Megrez2-3x7B-A3B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Infinigence/Megrez2-3x7B-A3B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Infinigence/Megrez2-3x7B-A3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Infinigence/Megrez2-3x7B-A3B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Infinigence/Megrez2-3x7B-A3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Infinigence/Megrez2-3x7B-A3B with Docker Model Runner:
```
docker model run hf.co/Infinigence/Megrez2-3x7B-A3B
```

Megrez2-3x7B-A3B / README.md

liyadong

add model file and update readme

b20e44e 11 months ago

preview code

Raw

History Blame

8.02 kB

	---
	license: apache-2.0
	language:
	- en
	- zh
	pipeline_tag: text-generation
	tags:
	- moe
	- conversational
	library_name: transformers
	---

	<div align="center">
	<img src="./assets/megrez-logo.png" alt="Megrez Logo" width="400" />

	<br>
	<h1> Megrez2-3x7B-A3B </h1>

	<a href="https://github.com/infinigence/Infini-Megrez">
	<b>🔗 Github</b>
	</a>  \|
	<a href="https://github.com/infinigence/Infini-Megrez/blob/main/docs/tech_report.pdf">
	<b>📄 Tech Report</b>
	</a>  \|
	<a href="https://huggingface.co/spaces/Infinigence/Megrez2-3x7B-A3B">
	<b>💻 Demo</b>
	</a>  \|
	<a href="https://huggingface.co/Infinigence/Megrez2-3x7B-A3B/blob/main/assets/wechat-official.jpg">
	<b>💬 WeChat Official</b>
	</a>

	<br>

	<strong>[中文](https://huggingface.co/Infinigence/Megrez2-3x7B-A3B/blob/main/README_ZH.md) \| English</strong>

	</div>

	## Introduction

	Megrez2-3x7B-A3B is a device native large language model. Megrez2 takes advantages of both the accuracy of Mixture-of-Experts (MoE) architecture and the compact size of Dense models. This release model was trained on 8T Tokens of data. In the future, we plan to improve the model's reasoning and agent capabilities.

	## Model Card

	<div align="center">

	\| \| \|
	\|:---:\|:---:\|
	\| Architecture \| Mixture-of-Experts (MoE) \|
	\| Total Parameters \| 3x7B \|
	\| Activated Parameters \| 3B \|
	\| Experts Shared Frequency\| 3 \|
	\| Number of Layers (Dense layer included) \| 31 \|
	\| Number of Dense Layers \| 1 \|
	\| Attention Hidden Dimension \| 2048 \|
	\| MoE Hidden Dimension (per Expert) \| 1408 \|
	\| Number of Attention Heads \| 16 \|
	\| Number of Experts \| 64 \|
	\| Selected Experts per Token \| 6 \|
	\| Number of Shared Experts \| 4 \|
	\| Vocabulary Size \| 128,880 \|
	\| Context Length \| 32K \|
	\| Base Frequency of RoPE \| 5,000,000 \|
	\| Attention Mechanism \| GQA \|
	\| Activation Function \| SwiGLU \|
	</div>

	## Performance

	We evaluated Megrez2-3x7B-A3B using the open-source evaluation tool [OpenCompass](https://github.com/open-compass/opencompass) on several important benchmarks. Some of the evaluation results are shown in the table below.

	<div align="center">
	<table>
	<thead>
	<tr>
	<th align="center">Benchmark</th>
	<th align="center">Metric</th>
	<th align="center"><sup>Megrez2-3x7B<br>-A3B</sup></th>
	<th align="center"><sup>Megrez2-3x7B<br>-A3B-Preview</sup></th>
	<th align="center"><sup>SmallThinker-21B<br>-A3B-Instruct</sup></th>
	<th align="center"><sup>Qwen3-30B-A3B</sup></th>
	<th align="center"><sup>Qwen3-8B</sup></th>
	<th align="center"><sup>Qwen3-4B<br>-Instruct-2507</sup></th>
	<th align="center"><sup>Phi4-14B<br>(nothink)</sup></th>
	<th align="center"><sup>Gemma3-12B</sup></th>
	</tr>
	</thead>
	<tbody>
	<tr>
	<td align="center">Activate Params (B)</td>
	<td align="center"></td>
	<td align="center">3.0</td>
	<td align="center">3.0</td>
	<td align="center">3.0</td>
	<td align="center">3.3</td>
	<td align="center">8.2</td>
	<td align="center">4.0</td>
	<td align="center">14.7</td>
	<td align="center">12.2</td>
	</tr>
	<tr>
	<td align="center">Stored Params (B)</td>
	<td align="center"></td>
	<td align="center">7.5</td>
	<td align="center">7.5</td>
	<td align="center">21.5</td>
	<td align="center">30.5</td>
	<td align="center">8.2</td>
	<td align="center">4.0</td>
	<td align="center">14.7</td>
	<td align="center">12.2</td>
	</tr>
	<tr>
	<td align="center">MMLU</td>
	<td align="center">EM</td>
	<td align="center">85.4</td>
	<td align="center"><strong>87.5</strong></td>
	<td align="center">84.4</td>
	<td align="center">85.1</td>
	<td align="center">81.8</td>
	<td align="center">-</td>
	<td align="center">84.6</td>
	<td align="center">78.5</td>
	</tr>
	<tr>
	<td align="center">GPQA</td>
	<td align="center">EM</td>
	<td align="center"><strong>58.8</strong></td>
	<td align="center">28.8</td>
	<td align="center">55.0</td>
	<td align="center">44.4</td>
	<td align="center">38.9</td>
	<td align="center">62</td>
	<td align="center">55.5</td>
	<td align="center">34.9</td>
	</tr>
	<tr>
	<td align="center">IFEval</td>
	<td align="center">Prompt<br>Strict</td>
	<td align="center"><strong>87.7</strong></td>
	<td align="center">80.2</td>
	<td align="center">85.8</td>
	<td align="center">84.3</td>
	<td align="center">83.9</td>
	<td align="center">83.4</td>
	<td align="center">63.2</td>
	<td align="center">74.7</td>
	</tr>
	<tr>
	<td align="center">MATH-500</td>
	<td align="center">EM</td>
	<td align="center"><strong>87.2</strong></td>
	<td align="center">81.6</td>
	<td align="center">82.4</td>
	<td align="center">84.4</td>
	<td align="center">81.6</td>
	<td align="center">-</td>
	<td align="center">80.2</td>
	<td align="center">82.4</td>
	</tr>
	</tbody>
	</table>
	</div>

	## How to Run

	### Transformers

	The latest version of `transformers` is recommended or `transformers>=4.52.4` is required.
	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 = "Infinigence/Megrez2-3x7B-A3B"
	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": "世界上最高的山峰是哪座？"},
	]
	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)

	# 世界上最高的山峰是珠穆朗玛峰（Mount Everest），位于喜马拉雅山脉的中尼边境。珠穆朗玛峰的海拔高度为8,848.86米（29,031.7英尺），这一数据是由中国和尼泊尔在2020年共同宣布的最新测量结果。珠穆朗玛峰不仅是登山爱好者的圣地，也是地理和科学研究的重要对象。
	```

	### 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
	```

	### llama.cpp

	Coming soon...

	## How to Deploy

	Megrez2-3x7B-A3B support using `vLLM` and `SGLang` as inference backends. For more information, please visit the [gitHub repository](https://github.com/infinigence/Infini-Megrez).

	## Best Practice

	To achieve optimal performance, we recommend the following settings:

	1. Sampling Parameters: we suggest using Temperature=0.7 and TopP=0.9 .

	2. Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking.
	* Math Problems: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
	* Multiple-Choice Questions: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the answer field with only the choice letter, e.g., "answer": "C"."

	## License Agreement

	All our open-weight models are licensed under Apache 2.0.

	## Citation

	If you find our work helpful, feel free to give us a cite.

	```bibtex
	@misc{li2025megrez2technicalreport,
	title={Megrez2 Technical Report},
	author={Boxun Li and Yadong Li and Zhiyuan Li and Congyi Liu and Weilin Liu and Guowei Niu and Zheyue Tan and Haiyang Xu and Zhuyu Yao and Tao Yuan and Dong Zhou and Yueqing Zhuang and Bo Zhao and Guohao Dai and Yu Wang},
	year={2025},
	eprint={2507.17728},
	archivePrefix={arXiv},
	primaryClass={cs.CL},
	url={https://arxiv.org/abs/2507.17728},
	}
	```

	## Contact

	If you have any questions, please feel free to submit a GitHub issue or contact [WeChat groups](https://huggingface.co/Infinigence/Megrez2-3x7B-A3B/blob/main/assets/wechat-group.jpg).