MiniMax M2.7 launches self-evolving agent model

MiniMax is making a bigger bet than a benchmark bump: M2.7 is positioned as a model-plus-harness stack that learns how to work better at agentic tasks. That is compelling, but the self-evolving claim mostly describes smarter scaffolding around the model, which is still a meaningful step for builders. Officially, M2.7 can build agent harnesses, use Agent Teams, search tools dynamically, and handle coding, debugging, research, and office-document workflows. MiniMax reports 56.22% on SWE-Pro, 55.6% on VIBE-Pro, and a 1495 GDPval-AA ELO, putting the release in serious frontier-agent territory. The interesting part is the loop: memory, self-feedback, evals, and scaffold edits over 100 rounds. Shipping it via API plus MiniMax Agent, with open-source weights, makes it useful both for internal teams and for builders who want a cheaper agentic model to slot into workflows. If the benchmarks hold up outside MiniMax’s own harness, this could be one of the cleaner examples of model as worker instead of model as chatbot.