MiMo-V2-Pro

MiMo-V2-Pro is Xiaomi's flagship foundation model purpose-built for real-world agentic workloads, serving as the decision-making layer that orchestrates complex workflows and drives production engineering tasks. Xiaomi describes the model as expanding the action space of frontier intelligence, generalizing from coding to broader agent operations. It is positioned as the immediate predecessor to MiMo-V2.5, which builds on the same backbone while adding native multimodality.

On agent-focused benchmarks relevant to deployment, MiMo-V2-Pro posts a General Agent Tool Use score of 61.5, ClawEval score of 81.0, and a PinchBench average of 1426, placing it ahead of several frontier peers on tool use and claw-style evaluations. The MiMo-V2.5 release further notes that MiMo-V2-Pro's language backbone derives from MiMo-V2-Flash's hybrid sliding-window attention architecture, a design later extended with visual and audio encoders. For practitioners, the model fits well in agent stacks that need a text-only reasoning controller with structured output, tool calling, and temperature control to coordinate downstream systems, while teams that require native visual or audio understanding may look to the follow-up release instead.