On March 18th, MiniMax introduced its latest Agent flagship large-scale model, M2.7, debuting its prowess as a "self-evolving model." Through the implementation of an Agent Harness system, this model actively engages in its own training and refinement processes. In certain research and development scenarios, it autonomously manages 30% to 50% of the workload, resulting in an approximate 30% enhancement in internal evaluation outcomes. M2.7 demonstrated an impressive 56.22% accuracy rate in the SWE-bench Pro test. It supports end-to-end project execution and offers a nuanced understanding of intricate systems. Additionally, it attained a top score of 1495 in the ELO rating for office scenarios, surpassing all other open-source models. This cutting-edge model is now accessible worldwide.
