MiniMax M2.7 helped build its own training harness
MiniMax says its new M2.7 model helped run its own reinforcement-learning experiments and improved its own agent scaffold, while posting self-reported coding scores close to Anthropic's Opus.
MiniMax has released M2.7, the latest model in its M2 series. The company calls it its "first model deeply participating in its own evolution." All benchmark figures below come from MiniMax and have not been independently verified.
On software engineering, MiniMax reports 56.22% on SWE-Pro, which it describes as "nearly approaching Opus's best level" without naming an Opus version. It also reports 55.6% on VIBE-Pro, which tests end-to-end project delivery, and 57.0% on Terminal Bench 2. On office work, M2.7 scored an Elo of 1495 on GDPval-AA, which MiniMax says is the highest of any open-source model. The company also says the model kept a 97% skill adherence rate while working with more than 40 complex skills, each over 2,000 tokens.
The "self-evolution" claims are more concrete than the label. An internal version of M2.7 built a research agent harness for MiniMax's RL team. The harness reviews literature, launches experiments, reads logs, debugs, and opens merge requests. MiniMax says the model handles 30–50% of that workflow, and humans step in for key decisions.
In one test, M2.7 worked on its own for more than 100 rounds to improve a model's coding performance on an internal scaffold. In each round it analysed failures, changed code, ran evaluations, and kept or reverted the change. Its changes included tuning sampling parameters and adding loop detection, and MiniMax reports a 30% gain on internal evaluation sets.
MiniMax also entered M2.7 in 22 machine-learning competitions from OpenAI's MLE Bench Lite, each runnable on a single A30 GPU. The source text cuts off before the detailed results.