MiniMax: MiniMax M1

MiniMax M1 is built around a hybrid Mixture-of-Experts architecture that houses 456 billion total parameters, with 45.9 billion activated per token during inference. This design is paired with a proprietary lightning attention mechanism that lets the model process very long sequences while keeping computational costs manageable. The model is specifically engineered for extended reasoning scenarios—complex, multi-step tasks that demand sustained context across software engineering problems, agentic tool use, and mathematical challenges. Its architecture reflects a deliberate trade-off: the large parameter count supports deep capability across diverse reasoning domains, while the MoE sparsity helps maintain competitive FLOP efficiency despite the model's scale.

The model is trained using a custom reinforcement learning pipeline called CISPO, which shapes its reasoning behavior through iterative feedback rather than traditional supervised fine-tuning alone. This training approach contributes to the model's strong showing across benchmarks including FullStackBench, SWE-bench, MATH, GPQA, and TAU-Bench, where it often outperforms comparable open models such as DeepSeek R1 and Qwen3-235B. MiniMax M1's advantages become clear in speed and intelligence benchmarks, particularly in coding and mathematical tasks, and it ships with a dedicated reasoning mode for chain-of-thought problems. As an open-weight model, it invites deployment and experimentation across research and production environments that value transparency and extensibility.