MiMo-V2-Pro

MiMo-V2-Pro is positioned as Xiaomi's flagship foundation model, designed to act as the brain of agent systems rather than a general conversational assistant. With over one trillion total parameters and a one-million-token that quick-info value window, it is built to orchestrate complex workflows, drive production engineering tasks, and reliably chain together tools. The model pushes what Xiaomi calls an expanded "action space" for intelligence, moving beyond code generation into the autonomous operation of digital "claws," and it integrates with general agent frameworks like OpenClaw. Its design intent is clearly agentic: long that quick-info value for sustained reasoning, tool calling for real-world execution, and temperature control for tunable behavior in production pipelines. The successor MiMo-V2.5-Pro further reveals the underlying architecture as a Mixture-of-Experts design with 42B active parameters on a hybrid-attention backbone, suggesting that the V2-Pro family trades raw dense compute for sparse, routed reasoning at scale.

MiMo-V2-Pro emerged from a team led by Fuli Luo, a veteran of the DeepSeek R1 effort, and was released alongside a broader Xiaomi model lineup. On agent-focused evaluations it lands in the global top tier, scoring 61.5 on General Agent Tool Use and 81.0 on ClawEval, with a PinchBench average of 1426, placing its perceived performance near that of Claude Opus 4.6 and ahead of several established frontier systems. On software engineering it records 55.0 on Coding Agent and 71.5 on SWE-Bench Pro, indicating solid but not yet top-of-leaderboard coding ability that the V2.5-Pro follow-up later extends to 57.2 and 73.7 respectively. The model is offered at a fraction of the cost of comparable Western frontier APIs, and Xiaomi has signaled plans to open-source a variant once the weights are stable. For practitioners, MiMo-V2-Pro is best understood as a long-context, tool-oriented reasoning engine suited to multi-step agent workflows, complex engineering tasks, and any setting where expansive that quick-info value and reliable tool orchestration matter more than open-weight availability.