MiniMax Releases Open-Source M2.7 AI Model with Self-Evolution Capabilities

Chinese AI company MiniMax has released its new open-source model, MiniMax M2.7, on Hugging Face, according to an announcement on X. The model contains 229 billion parameters and achieves state-of-the-art results on SWE-Pro (56.22%) and Terminal Bench 2 (57.0%).

MiniMax M2.7 is built using a Mixture-of-Experts architecture and is designed for multi-agent collaboration. It supports a range of professional tasks including software engineering, office productivity, and financial analysis. The model demonstrated strong results in benchmarks such as SWE Multilingual (76.5) and Multi SWE Bench (52.7%), and achieved an ELO score of 1495 on GDPval-AA, ranking highest among open-weight models.

The model introduces a self-evolution mechanism that allows it to autonomously analyze errors, modify its own code, and test improvements. During internal trials, M2.7 ran over 100 autonomous optimization cycles, improving its performance by 30%. It also reduced production debugging times to under three minutes in real-world tests.

MiniMax M2.7 is available for download and deployment on Hugging Face, with additional documentation and usage guides provided by the company.