MiniMax-M1

MiniMax-M1 is the world's first open-weight, large-scale hybrid-attention reasoning model, released by Shanghai-based MiniMax in June 2025. It pairs a 456-billion-parameter Mixture-of-Experts backbone (45.9B active per token) with Lightning Attention, natively supporting a 1-million-token context window, eight times that of DeepSeek-R1.

The model was trained with large-scale reinforcement learning using MiniMax's CISPO algorithm, with the full RL run completing in three weeks on 512 H800 GPUs for a reported rental cost of about $535,000. Two variants ship with 40K and 80K thinking budgets, and benchmarks place it at or above strong open-weight baselines on software engineering, tool use, and long-context tasks.