"M2.7 is our first model which deeply participated in its own evolution." MiniMax's new model ran 100+ autonomous cycles during development, analyzing its own failures, rewriting its own code, running evals, deciding what to keep. The company also said It handled 30-50% of its own training workflow with no human in the loop. The result: Competitive performance with Opus 4.6, GPT 5.4, and Gemini 3.1 Pro on major benchmarks. The best models of the future will all build themselves