MiniMax M2.7 Ships: Self-Evolving Model Ran 30–50% of Its Own RL Training Workflow

MiniMax·high signal

MiniMax released M2.7, a model that autonomously participated in 30–50% of its own reinforcement learning development — building skills, reading logs, debugging, and analyzing metrics without human direction. On SWE-Pro it scores 56.22% and achieves 97% skill adherence across 40+ complex skills each exceeding 2,000 tokens. Available on OpenRouter at $0.30/$1.20 per million tokens with 204K context, making it a cost-competitive option for agent pipelines.