The MiniMax M2.7 model has been unveiled, with its weights now available on Hugging Face, positioning it as a formidable contender among leading closed models. Performance metrics include a 56.22% score on SWE-Pro (a software engineering benchmark), closely rivaling Claude Opus 4.6; a 57.0% result on Terminal Bench 2; and an ELO rating of 1495 on GDPval-AA, which assesses real-world knowledge work tasks across various jobs. To put this into context, it’s the highest-scoring open-weight model, slightly trailing behind Opus 4.6, Sonnet 4.6, and GPT-5.4.
This innovative 230B-parameter Mixture of Experts model utilizes only 10B active parameters per inference pass, offering top-tier output without the need for equivalent computational resources. MiniMax claims that it was the first to autonomously participate in its own development; an internal version executed over 100 rounds of self-optimization, restructured its framework, and improved by 30% without human intervention.
However, soon after releasing the weights, the Chinese AI lab MiniMax quietly revised the licensing terms. Now, commercial use requires written approval from MiniMax, while non-commercial applications remain free and unrestricted, including for research and personal projects.
The shift drew immediate attention on platforms like Hacker News and Hugging Face, where developers expressed concerns. The main issue is the license being labeled as