MiniMax‑M1Open‑source large‑scale reasoning model with 1 million token context and hybrid Mixture‑of‑Experts architecture.

MiniMax-M1 is an open-weight, large-scale hybrid-attention reasoning model. It's powered by a hybrid Mixture-of-Experts (MoE) architecture combined with a lightning attention mechanism, enabling efficient scaling of test-time compute. The model natively supports a context length of 1 million tokens and is trained using large-scale reinforcement learning (RL) on diverse problems. It outperforms other strong open-weight models on complex software engineering, tool using, and long context tasks. Experiments on standard benchmarks show that MiniMax-M1 outperforms other models in category tasks such as mathematics, coding, software engineering, agentic tool use, and long-context understanding. The model is particularly suitable for complex tasks that require processing long inputs and thinking extensively. MiniMax-M1 serves as a strong foundation for next-generation language model agents to reason and tackle real-world challenges. The benchmark performance comparison of leading commercial and open-weight models across different category tasks highlights the model's performance. The technical report provides more information about the model's architecture, training protocol, and evaluation results.