MiniMax M2.7 is a reasoning-focused language model engineered specifically for agentic workflows, where it serves as the underlying engine powering popular developer tools like Claude Code, Kilo Code, and OpenClaw. What sets M2.7 apart is its self-evolving design philosophy—rather than relying solely on human-led fine-tuning, the model actively participates in building, monitoring, and optimizing its own reinforcement learning harnesses. This recursive self-improvement approach marks a shift toward models that act as architects of their own progress, not just products of human research. The model's strengths center on software engineering, demonstrating 56.22% on SWE-Pro benchmarks that match GPT-5.3-Codex, along with strong performance on multilingual code evaluation and end-to-end project delivery across web, mobile, and desktop platforms.

The training lineage of M2.7 reflects a novel post-training methodology where an internal version of the model autonomously ran over 100 optimization rounds—analyzing failure trajectories, modifying code, evaluating results, and deciding whether to keep or revert changes. This self-directed reinforcement learning loop delivered a 30% improvement on internal programming benchmarks. On MLE Bench Lite, M2.7 achieved a 66.6% medal rate, ranking second only behind Opus-4.6 and GPT-5.4 across 22 machine learning competitions. The model also introduces native agent teams with role identity and autonomous decision-making across complex state machines, achieving 97% skill compliance across more than 40 complex skills. These capabilities make M2.7 particularly well-suited for production-scale software engineering with long-horizon agentic execution and stable multi-agent collaboration.