MiniMax-M1

MiniMax-M1 was the first open-weight large-scale hybrid-attention reasoning model, released by MiniMax in June 2025 under Apache-2.0. The 456B-parameter mixture-of-experts activates 45.9B parameters per token and interleaves lightning attention, a linear-attention variant, with periodic softmax attention blocks, which is what lets it natively handle a 1 million token context while using roughly a quarter of DeepSeek R1's FLOPs at 100K-token generation lengths. Post-training used large-scale reinforcement learning with CISPO, an algorithm that clips importance sampling weights instead of token updates, and the model ships in two versions with 40K and 80K thinking budgets. Benchmark strengths center on long-context understanding, mathematical reasoning, tool use, and software engineering, including 56.0 percent on SWE-bench Verified. MiniMax recommends vLLM 0.9.2 or newer for serving and provides Transformers deployment and function-calling guides, with weights on Hugging Face; hosted access runs through the MiniMax API platform and chatbot. Hardware needs are datacenter-class given the parameter count.