MiniMax released M2.7 on March 18, a 230B sparse MoE model (10B parameters active per token) that the company describes as "the first model deeply participating in its own evolution." Over 100 iterative rounds the model analyzed its own failures, rewrote its training code, ran evaluations, and decided what to revert — yielding a claimed 30% performance improvement. M2.7 scores 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, targets long-horizon agentic workflows, and is available on NVIDIA's platform and HuggingFace under open weights.