MiniMax-M1
lastmod 2025-06-19created 2025-06-191 edits
An open-weight, large-scale hybrid-attention reasoning model
About
- MiniMax-M1 is the world's first open-weight, large-scale hybrid-attention reasoning model
- Excels at long-context processing thanks to a hybrid MoE architecture with 456 billion parameters and a lightning attention mechanism
- Can solve a wide range of problems efficiently by adopting RL-based training and the CISPO algorithm
- On benchmarks, shows excellent performance compared with existing models such as DeepSeek-R1 and Qwen3-235B in complex software engineering, tool use, long inputs, and more
- Offers various inference environments and supporting tools, an API, and a chatbot, making it highly valuable as a foundation for next-generation language model agents