MiniMax releases open-weight MiniMax-M1 hybrid-attention reasoning model

MiniMax has released MiniMax-M1, the world's first open-weight large-scale hybrid-attention reasoning model featuring a Mixture-of-Experts architecture and lightning attention mechanism. The model natively supports a context length of 1 million tokens and is trained using reinforcement learning on diverse problems including software engineering.

• MiniMax-M1 contains 456 billion parameters with 45.9 billion activated per token, offering 8x the context size of DeepSeek R1.
• The model consumes only 25% of the FLOPs compared to DeepSeek R1 at a generation length of 100K tokens.
• Two versions are available with 40K and 80K thinking budgets, with the M1-80k variant outperforming models like Qwen3-235B on complex tasks.
• The release includes usage instructions for libraries such as Transformers, vLLM, and SGLang, alongside Docker images for local deployment.

MiniMax-M1 serves as a foundation for next-generation language model agents to reason and tackle real-world challenges through efficient scaling of test-time compute.