Chinese AI startup MiniMax has released its M2.7 model, a proprietary large language model designed to power AI agents and integrate with third-party tools like Claude Code and Kilo Code. The model represents a significant technical achievement in that it can autonomously build, monitor, and optimize its own reinforcement learning systems.
Unlike previous models that relied solely on human-led fine-tuning, M2.7 can perform 30-50% of reinforcement learning research workflows independently. The model is categorized as a reasoning-only text system that delivers intelligence comparable to leading competitors while maintaining significantly higher cost efficiency, according to MiniMax.