Chinese AI startup MiniMax has released its new proprietary large language model, MiniMax M2.7, which can autonomously manage 30 to 50 percent of its own reinforcement learning research workflow. The model, designed for powering AI agents and third-party tools such as Claude Code and OpenClaw, marks a major step toward self-improving AI systems. MiniMax reports that M2.7 can autonomously debug, analyze metrics, and optimize its own code through iterative loops, achieving a 66.6 percent medal rate in machine learning competitions and matching performance levels of leading global models.

Compared to its predecessor M2.5, M2.7 shows significant improvements in software engineering, professional office tasks, and hallucination reduction. It matches top-tier benchmarks like GPT-5.3-Codex while maintaining one of the lowest operational costs among frontier AI models. The model is available through the MiniMax API and integrates with over 11 major developer tools, including Cursor, Zed, and Kilo Code.

MiniMax’s move toward proprietary models aligns with a broader industry trend among Chinese AI firms shifting from open-source to closed systems. The company positions M2.7 as a cost-efficient, production-ready model for enterprises seeking AI-driven automation and self-optimizing agent workflows.