MiniMax M2.7 debuts self-improving agents

The big story here isn’t just another benchmark bump; MiniMax is treating the model, the harness, and the eval loop as one product. That’s exactly where agentic coding models are headed.

• –The 100+ self-optimization cycles suggest eval design and scaffold quality are becoming part of the moat, not just supporting tooling.
• –For developers, the real question is whether those gains survive messy repos, flaky tools, and rollback-heavy workflows, not just clean internal tests.
• –If the reported improvements translate, M2.7 could be a serious value play for coding agents and internal copilots.
• –MiniMax keeps leaning into practical productivity use cases, which is where model competition is getting fiercest right now.