MiniMax M2.7

MiniMax M2.7 is designed as an agent-centric model that excels at managing complex, multi-step productivity tasks. Its architecture is built to support sophisticated agent harnesses, allowing it to utilize dynamic tool search and coordinate within agent teams to solve elaborate problems. By focusing on professional software engineering and technical workflows, the model is engineered to handle intricate coding environments and terminal-based operations, making it a robust choice for developers and organizations requiring high-level automation and autonomous problem-solving capabilities.

The model features a unique development lineage defined by a self-evolution cycle, where it actively participated in its own refinement. During its training, the model autonomously updated its internal memory, constructed dozens of specialized skills for reinforcement learning experiments, and iteratively improved its own learning process. This self-optimization approach allowed an internal version to refine a programming scaffold over 100 rounds of trial and error, resulting in significant performance gains. With strong results on benchmarks like MLE Bench Lite and SWE-Pro, M2.7 represents a forward-looking shift toward AI systems that can analyze their own failure trajectories and adapt their internal logic to meet demanding real-world requirements.