MiniMax M2.7 targets complex agent workflows

MiniMax has introduced M2.7, a large language model focused on complex agent development and productivity tasks. The company highlights gains in real-world software engineering, including end-to-end project delivery, log analysis for bug hunting, code security and machine learning work.

The model also targets professional office workflows, with improved domain expertise and task delivery across Excel, PPT and Word. On GDPval-AA, M2.7 achieves an ELO score of 1495, described by MiniMax as the highest among open-source models, alongside stronger multi-turn editing and high-fidelity document modification.

MiniMax says M2.7 can operate in complex environments, maintaining a 97% skill adherence rate across 40 complex skills (>2000 Token) cases. On the SWE-Pro benchmark, M2.7 scores 56.22%, with related results of VIBE-Pro 55.6% for full project delivery and Terminal Bench 2 at 57.0% for complex engineering system understanding.

Developer access includes quick API integration, standard M2.7 and M2.7-highspeed versions with identical results, automatic Cache support and integration with AI coding tools. Token Plan users receive higher inference speeds, and MiniMax says its M2.7-based general Agent platform is now fully open.