Chinese AI lab MiniMax released M2.7 on July 27, a model whose defining feature is not raw benchmark performance but verified self-improvement. In internal testing, M2.7 ran entirely autonomously through more than 100 optimization rounds, systematically tuning its own behavior and achieving a 30% performance gain on internal evaluation sets without human retraining or intervention.
How the Self-Optimization Works
According to MiniMax’s technical announcement, the self-evolution process follows a concrete loop: analyze failure trajectories, plan changes, modify scaffold code, run evaluations, compare results, then decide whether to keep or revert each change. Over 100+ rounds, M2.7 discovered several optimizations on its own. It systematically searched for optimal combinations of sampling parameters (temperature, frequency penalty, presence penalty). It designed specific workflow guidelines, such as automatically searching for the same bug patterns in other files after applying a fix. It added loop detection to its own agent scaffold.
MiniMax describes the capability as “the first model deeply participating in its own evolution.”
Benchmark Context
On standard benchmarks, M2.7 scores 56.22% on SWE-Pro (approaching Claude Opus levels, per MiniMax), 55.6% on VIBE-Pro for end-to-end project delivery, and 57.0% on Terminal Bench 2 for complex engineering system comprehension. Its GDPval-AA ELO score of 1,495 is the highest among open-source models for office productivity tasks.
In a separate test, MiniMax had M2.7 compete in 22 machine learning competitions from OpenAI’s MLE Bench Lite, each running on a single A30 GPU. The model was given 24 hours per trial to iteratively improve its own solutions. Three runs yielded an average medal rate of 66.6%, trailing only Claude Opus 4.6 (75.7%) and GPT-5.4 (71.2%), and matching Gemini 3.1 (66.6%), according to MiniMax’s results.
Internal Deployment at MiniMax
MiniMax is already using M2.7 to accelerate its own model development. The company describes a workflow where M2.7 handles 30% to 50% of the reinforcement learning research pipeline: literature review, experiment specification, data pipelining, experiment monitoring, log analysis, debugging, metric analysis, code fixes, merge requests, and smoke tests. Human researchers interact only for critical decisions.
The model’s harness also self-iterates. According to MiniMax, M2.7’s internal harness “autonomously collects feedback, builds evaluation sets for internal tasks, and based on this continuously iterates its own architecture, skills/MCP implementation, and memory mechanisms.”
The Governance Question
Self-optimizing agents move from theoretical concept to production capability with M2.7. The model can modify its own sampling parameters, rewrite its workflow logic, and improve its scaffold code, all without human approval per iteration. MiniMax frames this as the beginning of “full autonomy” in AI development, stating that “future AI self-evolution will gradually transition towards full autonomy, coordinating data construction, model training, inference architecture, evaluation, and other stages without human involvement.”
That framing raises a specific governance question: if an agent autonomously modifies its own behavior 100+ times, who audits the cumulative drift? Each individual change may be small, but the aggregate effect transforms the agent’s operating characteristics. The NVIDIA Open Secure AI Alliance, announced the same day, is building cryptographic audit trails and identity frameworks for exactly this class of problem.