World's First Open-Source
Hybrid Attention Reasoning Model

MiniMax-M1 is the world's first open-weight, large-scale hybrid-attention reasoning model. It features a unique hybrid Mixture-of-Experts (MoE) architecture combined with Lightning Attention mechanism. With 456B total parameters and only 45.9B activated per token, it offers superior efficiency while supporting 1M token context length natively, which is 8x larger than many competing models.

MiniMax-M1 is highly efficient thanks to its Lightning Attention mechanism. Compared to DeepSeek R1, M1 consumes only 25% of the FLOPs at a generation length of 100K tokens. This makes it particularly suitable for complex tasks requiring extensive reasoning and long input processing.

MiniMax-M1 comes in two versions: M1-40K and M1-80K, referring to their thinking budgets. The 80K version offers enhanced reasoning capabilities for more complex tasks, while the 40K version provides a good balance of performance and efficiency. Both versions share the same core architecture and 456B parameter count.

MiniMax-M1 demonstrates strong performance across various benchmarks. On AIME 2024, it achieves 86.0% (80K) and 83.3% (40K). For software engineering tasks like SWE-bench Verified, it scores 56.0% and 55.6% respectively. It particularly excels in long-context tasks, achieving 73.4% on OpenAI-MRCR (128k) and 56.2% on 1M token tasks.

MiniMax-M1 can be deployed using multiple methods: vLLM (recommended for production), Transformers, or HuggingFace. The model is available on HuggingFace Hub and GitHub. For general use, you can try the online chatbot at chat.minimaxi.com or use the API for development purposes.

Lightning Attention is MiniMax-M1's innovative attention mechanism that enables efficient scaling of test-time compute. It combines with traditional Softmax Attention in a hybrid architecture, allowing the model to process extremely long contexts (up to 4M tokens during inference) while maintaining computational efficiency.

Yes, MiniMax-M1 supports function calling capabilities. The model can identify when external functions need to be called and output function call parameters in a structured format. This makes it suitable for agentic applications and tool use scenarios.

MiniMax-M1 is trained using large-scale reinforcement learning (RL) on diverse problems ranging from traditional mathematical reasoning to sandbox-based, real-world software engineering environments. The training utilizes CISPO, a novel algorithm that clips importance sampling weights instead of token updates, which outperforms other competitive RL variants.

MiniMax-M1 is released under the Apache 2.0 license, making it fully open-source and free for both commercial and research use. This allows developers and researchers to freely use, modify, and distribute the model while maintaining attribution.

For technical support, questions, or feedback, you can contact the MiniMax team at [email protected]. You can also follow the project on GitHub for updates and community discussions.