MiniMax M2.7 touts self-evolving harness

The hype is real, but the scary part is mostly the framing. M2.7 looks like an agentic training-and-optimization pipeline, not a deployed endpoint that silently mutates in customer production.

• –The official post says M2.7 updated its own memory, built skills, and modified scaffold code over 100+ optimization rounds, which points to a self-improving harness around the model rather than a model that freely edits its shipped weights.
• –Reproducibility will require versioning more than weights: prompts, tools, memory state, sampling params, harness code, and frozen eval sets all matter if the system keeps adapting.
• –Benchmark selection gets shakier when the optimization loop can influence the benchmark outcome, so external holdout tests and snapshot-based evals become essential.
• –For open source, this pushes value toward publishing the full system, not just open weights. A model without the harness may look weaker than the same model inside the closed loop.
• –Governance shifts from “who approved the model?” to “who approved the harness changes, tool permissions, and rollback rules?” That’s the real production risk surface.