Summary

MiniMax-M3 is a native multimodal large language model featuring a 1 million token context window. It is designed for deep semantic fusion across text, image, and video, targeting advanced applications in coding and collaborative AI agents. The model offers significant performance improvements in long-context processing and agentic tasks.

How It Works

M3 employs native multimodal training from inception, enabling profound semantic integration of text, image, and video data. Its core innovation is MiniMax Sparse Attention (MSA), a high-performance sparse attention mechanism engineered for million-token contexts. MSA dramatically reduces attention computation and memory requirements compared to standard methods like GQA, while preserving model quality, leading to substantial speedups in prefill and decoding.

Quick Start & Requirements

• Installation: Download model weights using hf download MiniMaxAI/MiniMax-M3 --local-dir MiniMax-M3.
• Inference: Recommended frameworks include SGLang, vLLM, and Transformers. Refer to their respective documentation for integration.
• Parameters: For optimal performance, use temperature=1.0, top_p=0.95, top_k=40.
• Resources: Requires significant computational resources for local deployment.
• Links: Technical Report: arXiv:2606.13392, Hugging Face Papers.

Highlighted Details

• 1 million token context window.
• Approximately 428 billion total parameters, with ~23 billion activated.
• Native multimodal training for text, image, and video.
• MiniMax Sparse Attention (MSA) provides 9x prefill and 15x decode speedups at 1M context, reducing per-token compute by 1/20 compared to M2.
• Achieves frontier-level performance on long-horizon agentic benchmarks, excelling in coding and cowork capabilities.

Maintenance & Community

• Contact: Inquiries can be directed to model@minimax.io.

Licensing & Compatibility

• No specific license information is provided in the README. Users should verify licensing terms before adoption, especially for commercial use.

Limitations & Caveats

• The README does not detail specific limitations, unsupported platforms, or known bugs. Users should conduct thorough testing for their specific use cases.