MiniMax Open-Sources M2.7, a 230B AI Model That Optimized Its Own Training

Chinese AI lab MiniMax released M2.7 on April 12, 2026, a 230-billion-parameter Mixture-of-Experts model that was given active write access to its own memory, skills, and training infrastructure during development. Unlike previous AI systems trained passively, M2.7 operated within an autonomous agentic scaffold that allowed it to propose and implement changes to its own reinforcement learning setup. Through this process, the model independently discovered optimizations in sampling parameters, workflow guidelines, and loop-detection logic, resulting in a 30% improvement in training efficiency. The model's weights are publicly available on Hugging Face, and its benchmark scores on SWE-bench Pro are reported to be comparable to GPT-5.3-Codex. MiniMax describes M2.7 as the first production model to document a concrete self-improvement loop running inside an agentic training harness.