MiniMax has announced the release of its M2.7 model, a major advancement in the M2-series lineup. M2.7 is publicly available via the MiniMax Agent and the MiniMax API Platform. This release marks a shift toward models that participate in their own evolution, using agent harnesses and reinforcement learning to optimize their capabilities, including memory updates, skill development, and iterative self-improvement.
M2.7 supports complex workflows in software engineering, office productivity, and research environments. Notable technical details include multi-agent collaboration, high skill adherence rates (97% across 40+ complex skills), and performance scores near the industry’s best on benchmarks such as SWE-Pro (56.22%) and VIBE-Pro (55.6%).
Compared to predecessor models, M2.7 demonstrates improved multilingual programming, code security, end-to-end project delivery, and deep system-level understanding. It also outperforms most open-source models in professional office tasks, with an ELO score of 1495 on GDPval-AA.
M2.7 introduces capabilities like autonomous debugging, research agent harnesses, and the OpenRoom demo for entertainment and interactive experiences. Early user feedback highlights its ability to produce deliverables that integrate directly into professional workflows. MiniMax, the company behind this development, continues to accelerate its transition into an AI-native organization by deploying M2.7 internally, using the model to automate and optimize research and development processes with minimal human intervention.
