The model that tuned itself

MiniMax released M2.7, an open-weights coding model, in March, and buried in the announcement is the part worth slowing down for: a private copy of the model was put in the optimizer's seat over its own workings. For 100-plus rounds with no human in the loop, it read its own failure logs, rewrote the scaffolding that wraps the model at runtime, re-tested, and kept or reverted each change.

It touches the harness around the model, not the training or the weights — nothing here retrains anything.

What it found is more telling than the loop itself. Not exotic self-rewriting — the craft a careful engineer would recognize: a better combination of sampling settings, a rule that after fixing a bug it should sweep other files for the same pattern, a loop-detector so the agent stops spinning. MiniMax reports the run lifted its internal scores about 30 percent and titles the page, honestly, 'Early Echoes of Self-Evolution.'

Two cautions travel with that number. It touches the harness around the model, not the model's training or its weights — nothing here retrains anything. And it is MiniMax's own figure on MiniMax's own evals; on an independent vibe-coding benchmark the 'self-evolved' M2.7 actually placed lower than the M2.5 it replaced. A model grading its own homework is the oldest problem in the field, and self-improvement is exactly where it bites hardest.

Still, the direction is the signal. The idea of a system editing its own code to get better is old — Sakana's Darwin-Godel Machine did it on coding tasks a year earlier — but here it is a frontier model doing it to itself, shipped inside a cheap, downloadable commercial product rather than a lab demo. The open question is no longer whether a model can tune its own scaffolding, but who checks the report card when it does.