What makes MiniMax M2.7 fundamentally different from previous AI models?

M2.7 is the first AI model designed to actively participate in its own evolution. It autonomously builds and monitors its reinforcement learning harnesses, analyzes failure trajectories, plans and executes improvements, and completed 100+ autonomous iteration cycles — achieving a 30% internal benchmark performance gain without any human engineering intervention.

How did M2.7 demonstrate autonomous self-improvement in Kaggle competitions?

M2.7 competed in 22 MLE-Bench Lite machine learning competitions using a self-designed three-module agent system (short-term memory, self-feedback, self-optimization). Over 24 hours of autonomous iteration, it won 9 gold, 5 silver, and 1 bronze medals, raising its average medal rate from ~50% to 66.6% — ranking third among all tested models.

How does M2.7's software engineering performance compare to GPT-5.3 Codex and Claude Opus 4.6?

M2.7 scores 56.22% on SWE-Pro, essentially matching GPT-5.3 Codex (56.8%); and 55.6% on VIBE-Pro for end-to-end project delivery, nearly on par with Claude Opus 4.6. In production debugging scenarios, M2.7 has reduced incident recovery times to under three minutes.

— AI DAILY

Sources

新浪科技