MiniMax M2.7 debuts self-evolving agents
MiniMax has launched M2.7 as the newest text model in its lineup, positioning it as the first model deeply involved in improving its own harness, memory, skills, and reinforcement-learning loop. The company says it targets software engineering, office workflows, and complex multi-step agent tasks, with claimed gains on SWE-Pro, GDPval-AA, Terminal Bench 2, and MM Claw.
This feels less like a routine model bump and more like MiniMax trying to productize recursive improvement: the model is being used to help build the very harnesses that train and evaluate it. If those lab gains hold up in messy real workflows, M2.7 could become one of the more interesting coding-and-agent models in the market.
- –The defining story is self-improvement: MiniMax says M2.7 helped build complex skills, update memory, and optimize its own RL harness through 100+ autonomous iterations.
- –The release is aimed at practical developer pain, not just synthetic leaderboard wins: repo delivery, log analysis, code security, and high-fidelity office document editing.
- –MiniMax is also packaging the model for builders, with API and high-speed variants plus an unchanged pricing story for token-plan users.
- –Community reaction is cautiously bullish, but the real test will be whether the benchmark gains survive long, flaky tool chains and real production tasks.
DISCOVERED
25d ago
2026-03-18
PUBLISHED
25d ago
2026-03-18
RELEVANCE
AUTHOR
hedgehog0