MiniMax drops open-source M1 reasoning model with 1M context
MiniMax just open-sourced MiniMax-M1, a hybrid-attention reasoning model with a 1 million token context window and 80k-token reasoning output. Thanks to its Lightning Attention mechanism and CISPO RL algorithm, deep reasoning costs roughly 30% of DeepSeek R1's compute, with the entire RL phase trained on 512 H800s for just $535k. It rivals Gemini 2.5 Pro in long context while being completely free on the web.