MiniMax M2.7

MiniMax M2.7 represents a shift toward models that actively participate in their own development, designed from the ground up for complex agentic workflows. The model excels at building intricate agent harnesses and completing elaborate productivity tasks by orchestrating Agent Teams, dynamic tool search, and complex skills. Its architecture supports professional software engineering and office task execution, making it well-suited for demanding coding environments and agent-based applications.

During its development cycle, M2.7 engaged in genuine self-evolution: it updated its own memory, constructed numerous complex skills for reinforcement learning experiments, and iteratively improved its learning process based on experimental results. An internal version of the model autonomously refined a programming scaffold across more than 100 rounds of analysis, code modification, evaluation, and selective retention, delivering measurable performance gains. On MLE Bench Lite's 22 machine learning competitions, M2.7 achieved a 66.6% medal rate, ranking behind only Opus-4.6 and GPT-5.4. The model is available with an extensive context window and powers AI coding tools in platforms like Claude Code, reflecting its practical strength in real-world development workflows.