MiniMax M1 40K

MiniMax-M1 is an open-source, large-scale reasoning model that uses a hybrid-attention architecture for efficient long-context processing. It supports up to a 1 million token context window and 80,000-token reasoning output, matching Gemini 2.5 Pro’s scale while being highly cost-effective. Its Lightning Attention mechanism reduces compute requirements to about 30% of DeepSeek R1’s, and a new reinforcement learning algorithm, CISPO, doubles convergence speed compared to other RL methods. Trained on 512 H800s over three weeks, M1 achieves near state-of-the-art results across software engineering, long-context, and tool-use benchmarks, outperforming most open models and rivaling top closed systems.

Benchmark results

Benchmark Score Tags Source
AIME 2024 83.3% self-reported llm-stats link →
AIME 2025 74.6% self-reported llm-stats link →
GPQA 69.2% self-reported llm-stats link →
Humanity's Last Exam 7.2% self-reported llm-stats link →
LiveCodeBench 62.3% self-reported llm-stats link →
LongBench v2 61.0% self-reported llm-stats link →
MATH-500 96.0% self-reported llm-stats link →
MMLU-Pro 80.6% self-reported llm-stats link →
Multi-Challenge 44.7% self-reported llm-stats link →
OpenAI-MRCR: 2 needle 128k 76.1% self-reported llm-stats link →
OpenAI-MRCR: 2 needle 1M 58.6% self-reported llm-stats link →
SimpleQA 17.9% self-reported llm-stats link →
SWE-Bench Verified 55.6% self-reported llm-stats link →
TAU-bench Airline 60.0% self-reported llm-stats link →
TAU-bench Retail 67.8% self-reported llm-stats link →
ZebraLogic 80.1% self-reported llm-stats link →