MiniMax Releases M2.7 Flagship Model: Self-Evolution Capability, Excellent Performance in Code and Office Tasks

MiniMax Technology released its new generation Agent flagship large model M2.7, showcasing the path of model self-evolution for the first time. By building an Agent Harness system, the model can deeply participate in its own training and optimization, undertaking 30%-50% of the workload in some R&D scenarios and improving performance by about 30% in internal evaluations.

M2.7 excels in code tasks: it achieves a 56.22% accuracy rate in the SWE-Pro code test, matching GPT-5.3-Codex; its VIBE-Pro score of 55.6% in repo-level code generation is close to Opus4.6. In office scenarios, it ranks highest among open-source models with an ELO score of 1495 in GDPval-AA evaluation, and its complex editing capabilities for Office suite tools are significantly improved.

Additionally, MiniMax open-sourced the interactive system OpenRoom to explore AI-native interaction methods, and M2.7 is now fully available on the MiniMax Agent and open platform.