MiniMax Opens Source M2.7 Model, Achieving Breakthroughs in AI Self-Training and Complex Task Capabilities

MiniMax recently announced the open-source release of the M2.7 model, which can deeply participate in its own training and optimization process, build complex Agent frameworks to complete high-difficulty productivity tasks.

In software engineering scenarios, M2.7 performs excellently, covering end-to-end project delivery, log analysis, bug troubleshooting, etc. It scored 56.22% in the SWE-Pro benchmark, close to top model levels; 55.6% in VIBE-Pro full project delivery, and 57.0% in Terminal Bench 2 complex system understanding.

Additionally, M2.7 has self-evolution capabilities: it can automatically collect feedback, build evaluation sets to optimize its architecture and skills. For programming ability optimization, it can run over 100 cycles independently, achieving a 30% performance improvement. The model also has the OpenRoom interaction system, extending AI interaction to visual interfaces, supporting real-time scene feedback and high scalability.