MiniMax: MiniMax M1

MiniMax-M1 is a cutting-edge, open-weight reasoning model engineered for exceptional performance with extended contexts and highly efficient inference. It employs a hybrid Mixture-of-Experts (MoE) architecture combined with a custom "lightning attention" mechanism, enabling it to process sequences up to 1 million tokens while maintaining competitive FLOP efficiency. With 456 billion total parameters and 45.9B active per token, this variant is specifically optimized for complex, multi-step reasoning tasks. Trained using a custom reinforcement learning pipeline (CISPO), M1 demonstrates superior capabilities in long-context understanding, software engineering, agentic tool use, and mathematical reasoning. Benchmarks consistently show strong performance across FullStackBench, SWE-bench, MATH, GPQA, and TAU-Bench, frequently surpassing other open models such as DeepSeek R1 and Qwen3-235B. It supports functions and streaming, making it versatile for various applications. With a context window of 1000K tokens and a max output of 4K tokens, it offers extensive processing power. Pricing is competitive at $0.40/2.20 per 1M tokens (input/output) for PRO access.