MiniMax-M2.7 is the first model in the MiniMax family to deeply participate in its own evolution, designed from the ground up for complex agentic workflows and elaborate productivity tasks. The model can build its own agent harnesses, coordinate Agent Teams, manage complex Skills, and dynamically search for tools as needed. Built on the OpenClaw Agent Harness framework, it brings native support for self-evolving behavior where the model itself can update its own memory, construct new skills, and steer improvements to its own learning pipeline—setting it apart from static models that simply execute tasks.

MiniMax-M2.7's training involved a distinctive self-improvement loop: an internal version autonomously ran over 100 rounds of scaffold optimization, analyzing failure trajectories, modifying code, running evaluations, and making keep-or-revert decisions to achieve roughly 30% performance gains without human intervention. It also independently managed a significant portion of its own reinforcement learning research workflows. On MLE-Bench Lite across 22 ML competitions, M2.7 achieved a 66.6% medal rate, approaching the performance of much larger proprietary models. With a 10 billion parameter footprint, it delivers competitive results on hard coding and agent benchmarks while running faster and at a fraction of the cost, making it well-suited for developers who need powerful, cost-efficient AI assistance for professional software engineering and complex multi-step tasks.