MiniMax M2.7: Self-Evolving AI Model That Participated in Its Own Development
MiniMax has released M2.7, a self-evolving AI model that ran over 100 autonomous development cycles — analyzing failures, rewriting code, running evaluations, and deciding whether to keep or revert changes — all without human intervention. The model handled 30-50% of its own reinforcement learning research workflow.
During autonomous development, M2.7 discovered optimizations that human researchers hadn't anticipated, including systematic sampling parameter search, workflow-specific guidelines, and loop detection mechanisms. This resulted in a 30% performance improvement on internal evaluation sets.
M2.7 scored 56.22% on SWE-Bench Pro and maintained a 97% skill compliance rate across 40 complex skills. It natively supports Agent Teams for multi-agent collaboration with role boundaries, adversarial reasoning, and protocol adherence — capabilities that must be internalized during training rather than achieved through prompting.
The model was announced on March 18, 2026, and is trending on Product Hunt with 149 upvotes. Covered by VentureBeat.
Official announcement: https://www.minimax.io/news/minimax-m27-en
← 返回所有文章
During autonomous development, M2.7 discovered optimizations that human researchers hadn't anticipated, including systematic sampling parameter search, workflow-specific guidelines, and loop detection mechanisms. This resulted in a 30% performance improvement on internal evaluation sets.
M2.7 scored 56.22% on SWE-Bench Pro and maintained a 97% skill compliance rate across 40 complex skills. It natively supports Agent Teams for multi-agent collaboration with role boundaries, adversarial reasoning, and protocol adherence — capabilities that must be internalized during training rather than achieved through prompting.
The model was announced on March 18, 2026, and is trending on Product Hunt with 149 upvotes. Covered by VentureBeat.
Official announcement: https://www.minimax.io/news/minimax-m27-en
评论