MiniMax M2

MiniMax M2 is an open-source large language model by MiniMax, built for agents and coding tasks. It delivers state-of-the-art tool use, reasoning, and search performance while maintaining exceptional cost-efficiency and speed, priced at just 8% of Claude 3.5 Sonnet’s cost and running at nearly double its inference speed (≈100 TPS). Designed for end-to-end agentic workflows, it excels at long-chain tool calling across Shell, Browser, Python, and other MCP tools. While slightly behind top overseas models in programming, it ranks among the best domestic models and top five globally on the Artificial Analysis benchmark. M2 powers the MiniMax Agent platform, available in Lightning Mode for fast tasks and Pro Mode for complex multi-step reasoning, and its weights, API, and deployment guides are freely available on Hugging Face, vLLM, and SGLang.

Benchmark results

Benchmark Score Tags Source
AA-Index 61.0% self-reported llm-stats link →
AIME 2025 78.0% self-reported llm-stats link →
BrowseComp 44.0% self-reported llm-stats link →
BrowseComp-zh 48.5% self-reported llm-stats link →
GPQA 78.0% self-reported llm-stats link →
Humanity's Last Exam 12.5% self-reported llm-stats link →
IF 72.0% self-reported llm-stats link →
LiveCodeBench 83.0% self-reported llm-stats link →
MMLU-Pro 82.0% self-reported llm-stats link →
Multi-SWE-Bench 36.2% self-reported llm-stats link →
SciCode 36.0% self-reported llm-stats link →
SWE-bench Multilingual 56.5% self-reported llm-stats link →
SWE-Bench Verified 69.4% self-reported llm-stats link →
Tau-bench 77.2% self-reported llm-stats link →
Tau2 Telecom 87.0% self-reported llm-stats link →
Terminal-Bench 46.3% self-reported llm-stats link →