MiMo V2.5 Pro is positioned as Xiaomi's flagship open-weight agent foundation model, engineered for sustained coherent behavior on complex, multi-step tasks. It is built on a hybrid-attention Mixture-of-Experts architecture with roughly 1.02 trillion total parameters and around 42 billion active parameters per pass, giving it broad capacity while keeping per-token inference costs contained. The model extends its that quick-info value window to the cataloged API limit, allowing it to ingest large codebases, lengthy contracts, or extended research reports as a single input rather than splitting them into chunks, and it supports a maximum output length in the high tens of thousands of tokens for long generations.
In Xiaomi's own benchmark suite, MiMo V2.5 Pro posts a Coding Agent score of 57.2, a SWE-Bench Pro score of 73.7, and a MiMo Coding Bench score of 68.4, reflecting its focus on real software-engineering workflows and tool-using productivity. Xiaomi highlights its ability to remain coherent across nearly one thousand tool calls in long-horizon runs, its single-pass completion advantage on complex multi-step tasks, and its strength in fuzzy instruction following. These characteristics make it well suited to autonomous coding agents, large-codebase refactoring, and research-assistant workloads that require combining deep retrieval, tool orchestration, and extended reasoning within a single session.
