MiniMax releases next-gen AI model M2.7 with self-evolution capabilities

  • M2.7 shows significant improvement over M2.5 in OpenClaw usage, MiniMax said.
  • M2.7 can autonomously build complex agent harnesses to complete highly complex productivity tasks.
MiniMax releases next-gen AI model M2.7 with self-evolution capabilities
(Image credit: A screenshot from MiniMax's website.)

Chinese artificial intelligence company MiniMax has released its next-generation large model, M2.7, showcasing a self-evolution path for AI models for the first time.

The model can deeply participate in its own training and iterative optimization processes, taking on about 30% to 50% of the workload in certain research and development scenarios, according to a statement from MiniMax on Wednesday.

M2.7 is capable of autonomously building complex agent harnesses and completing highly complex productivity tasks based on capabilities such as agent teams, complex skills, and tool search tools.

The model demonstrated end-to-end project delivery capabilities in real-world software engineering tasks, extending to a deep understanding of complex engineering systems.

In the SWE-Pro benchmark test, M2.7 scored 56.22%, nearly matching the best performance of Anthropic's Claude Opus, MiniMax said.

In professional office work, M2.7 also demonstrated commercial application value. Its ELO score on GDPval-AA ranked just behind Claude Opus 4.6, Claude Sonnet 4.6, and OpenAI's GPT-5.4, while surpassing GPT-5.3.

The model can process complex office documents and maintains a high degree of instruction adherence during multi-turn interactions, according to the company.

Beyond breakthroughs in productivity tools, M2.7 has improved the stability of role-playing, opening up new possibilities for interactive entertainment scenarios.

MiniMax also built an evaluation set called MM-Claw based on common tasks in OpenClaw, covering personal learning planning, office document processing, information research, and code development and maintenance.

M2.7 reached a level close to Claude Sonnet 4.6 in this test, with an accuracy rate of 62.7%, it said.

In the use of OpenClaw, M2.7 showed a significant improvement compared to M2.5, approaching Sonnet 4.6 in the MM-Claw evaluation.

MiniMax M2.7 is now available on the MiniMax Agent and open platforms.

MiniMax shares jumped nearly 24% in Hong Kong trading on Monday, largely driven by a surge in model token usage tied to the viral OpenClaw ecosystem.
Mar 9, 2026
AI News Alert
Subscribe to receive email notifications immediately when new articles about AI are published.
AI
View more channels