MiniMax has introduced MiniMax-M1, the world's first open-weight large-scale hybrid-attention reasoning model. Built on its previous MiniMax-Text-01 architecture, it combines a Mixture-of-Experts (MoE) design with a lightning attention mechanism to support 1 million token context and efficient test-time compute scaling.
The model features 456 billion total parameters with 45.9 billion activated per token, consuming only 25% of the FLOPs of DeepSeek R1 at 100K generation length. It was trained using large-scale reinforcement learning and two new versions with 40K and 80K thinking budgets. Experiments show it outperforms models like DeepSeek-R1 and Qwen3-235B on complex software engineering, tool use, and long-context tasks.
MiniMax-M1 serves as a foundation for next-generation language model agents by enabling efficient reasoning over long inputs and extensive thinking processes.