MiniMax M2.7 Launched: First AI Model Deeply Engaged in Self-Iteration, Programming Capability Close to Opus Level

MiniMax recently released M2.7, described as the first AI model deeply involved in self-iteration. The team built a research Agent framework based on early M2 versions, enabling the model to autonomously perform reinforcement learning skill building, memory updates, and process optimization—achieving a 30% performance improvement after over 100 iteration cycles in internal tests.

In terms of performance, M2.7 achieved an average 66.6% medal rate in 22 machine learning tasks on MLE Bench Lite, on par with Gemini-3.1. For programming, it scored 56.22% on SWE-Pro (matching GPT-5.3-Codex) and 55.6% on VIBE-Pro (close to Opus 4.6). In office scenarios, its GDPval-AA ELO score was 1500, ranking top 4 among 45 models; OpenClaw evaluation showed MM-Claw's accuracy at 62.7% (near Sonnet 4.6) with a 97% compliance rate for complex skills.

Additionally, MiniMax open-sourced the Agent interaction system OpenRoom and fully launched M2.7 on MiniMax Agent and its open platform.