MiniMax M2.7 Launches — Self-Evolving Proprietary Model Handled 30-50% of Its Own RL Training Workflow

MiniMax·high signal

MiniMax released M2.7 on March 18, a proprietary (not open-source) model that participated in its own development, autonomously handling 30-50% of its reinforcement learning research pipeline. Benchmark highlights: SWE-Pro 56.22% matching GPT-5.3-Codex, GDPval-AA ELO 1495 highest among comparable models, MM Claw 62.7% approaching Sonnet 4.6, available on OpenRouter at $0.30/$1.20 per million tokens. Unlike M2/M2.1/M2.5 which were MIT-licensed open-weight releases, M2.7 is closed-weight despite shipping via the same community channels.