unsloth/MiniMax-M3

1---2pipeline_tag: image-text-to-text3license: other4license_name: minimax-community5license_link: LICENSE6library_name: transformers7tags:8- multimodal9- moe10- agent11- coding12- video13base_model:14- MiniMaxAI/MiniMax-M315---16 17<div align="center">18  <img width="60%" src="figures/logo.svg" alt="MiniMax">19</div>20<hr>21 22<div align="center" style="line-height: 1.4; font-size:16px; margin-top: 30px;">23  Join Our 24  <a href="https://platform.minimaxi.com/docs/faq/contact-us" target="_blank" style="font-size:17px; margin: 2px;">25    💬 WeChat26  </a> | 27  <a href="https://discord.com/invite/DPC4AHFCBw" target="_blank" style="font-size:17px; margin: 2px;">28    🧩 Discord29  </a> 30  community.31</div>32<div align="center" style="line-height: 1.2; font-size:16px;">33  <a href="https://agent.minimax.io/" target="_blank" style="display: inline-block; margin: 4px;">34    MiniMax Agent35  </a> | 36  <a href="https://platform.minimax.io/docs/guides/text-generation" target="_blank" style="display: inline-block; margin: 4px;">37    ⚡️ API38  </a> | 39  <a href="https://github.com/MiniMax-AI/cli" style="display: inline-block; margin: 4px;">40    CLI41  </a> |42  <a href="https://www.minimax.io" target="_blank" style="display: inline-block; margin: 4px;">43    MiniMax Website44  </a> 45</div>46<div align="center" style="line-height: 1.2; font-size:16px; margin-bottom: 30px;">47  <a href="https://huggingface.co/MiniMaxAI" target="_blank" style="margin: 2px;">48    🤗 Hugging Face 49  </a> | 50  <a href="https://github.com/MiniMax-AI/MiniMax-M3" target="_blank" style="margin: 2px;">51    🐙 GitHub52  </a> | 53  <a href="https://www.modelscope.cn/organization/MiniMax" target="_blank" style="margin: 2px;">54    🤖️ ModelScope55  </a> | 56  <a href="https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/main/LICENSE" style="margin: 2px;">57    📄 LICENSE58  </a>59</div>60 61MiniMax-M3 is a native multimodal model with 1M context. It has ~428B parameters and ~23B activated parameters.62 63**Highlights:**64- **Native Multimodality:** M3 undergoes mixed-modality training from the very first step, enabling deeper semantic fusion across text, image, and video.65- **Context Scaling via Sparse Attention:** M3 introduces MiniMax Sparse Attention (MSA) to improve long context efficiency. M3 delivers 9× prefill and 15× decode speedups compared to M2 at 1M context, reducing per-token compute to 1/20.66- **Coding & Cowork Capability:** M3 achieves frontier-level performance across long-horizon agentic benchmarks, excelling in both coding and cowork.67 68## Model Details69| | |70| --- | --- |71| Architecture | MoE + MSA (MiniMax Sparse Attention) |72| Total Parameters | ~428B |73| Activated Parameters | ~23B |74| Experts | 128 (4 active per token) |75| Layers | 60 |76| Context Length | 1M tokens |77| Modalities | Text, Image, Video |78| Precision | bfloat16 |79| Transformers | ≥ 4.52.4 (`trust_remote_code=True`) |80| License | [MiniMax Community License](LICENSE) |81<p align="center">82  <img width="100%" src="figures/benchmark.jpeg">83</p>84 85## How to Use86 87- [MiniMax Agent](https://agent.minimax.io/)88- [MiniMax API](https://platform.minimax.io/)89 90M3 supports two reasoning modes:91- **thinking** — for complex reasoning, agentic tasks, and long-horizon collaboration.92- **non-thinking** — for latency-sensitive scenarios such as chat and code completion.93 94## Local Deployment95 96Download the model:97 98```bash99hf download MiniMaxAI/MiniMax-M3 --local-dir MiniMax-M3100```101 102We recommend the following inference frameworks (listed alphabetically) to serve the model:103 104### SGLang105 106We recommend using [SGLang](https://docs.sglang.io/) to serve MiniMax-M3. Please refer to our [SGLang Deployment Guide](./docs/sglang_deploy_guide.md).107 108### vLLM109 110We recommend using [vLLM](https://github.com/vllm-project/vllm) to serve MiniMax-M3. Please refer to our [vLLM Deployment Guide](./docs/vllm_deploy_guide.md).111 112### Transformers113 114We recommend using [Transformers](https://github.com/huggingface/transformers) to serve MiniMax-M3. Please refer to our [Transformers Deployment Guide](./docs/transformers_deploy_guide.md).115 116### ModelScope117 118You can also get model weights from [ModelScope](https://modelscope.cn/models/MiniMax/MiniMax-M3).119 120### Inference Parameters121 122We recommend the following parameters for best performance: `temperature=1.0`, `top_p=0.95`, `top_k=40`. Default system prompt:123 124```125You are a helpful assistant. Your name is MiniMax-M3 and was built by MiniMax.126```127 128## Tool Calling Guide129 130Please refer to our [Tool Calling Guide](./docs/tool_calling_guide.md).131 132## Contact Us133 134Contact us at [model@minimax.io](mailto:model@minimax.io).