MiniMax introduced M2.7 on March 18, 2026 — the first model in its M-series to participate in its own evolutionary development cycle.
The defining feature is self-evolution. M2.7 ran entirely autonomously through an iterative loop: analyze failure trajectories, plan changes, modify scaffold code, run evaluations, compare results, and decide to keep or revert changes — for over 100 rounds. This process achieved a 30% performance improvement on internal evaluation sets without human intervention.
The model can build its own harness skills — reusable instruction sets up to 2,000+ tokens each — update its own memory store based on task outcomes, and run reinforcement learning experiments to optimize its own performance.
In MiniMax's internal trials, M2.7 handles 30-50% of the workflow in research environments, effectively serving as a semi-autonomous research assistant that can iterate on its own capabilities.
Benchmark performance is strong: M2.7 scored 56.22% on SWE-Pro, matching GPT-5.3-Codex levels. In document processing, it achieved an Elo score of 1495 on GDPval-AA, which MiniMax claims is the highest among open-source-accessible models.
Compared to M2.5 (released just a month earlier), M2.7 shows significant gains in software engineering and professional office tasks — improvements largely driven by the self-evolution process rather than traditional human-directed training.