MiMo-V2.5

MiMo-V2.5-Pro represents Xiaomi's flagship entry in the MiMo V2.5 lineup, replacing the earlier MiMo-V2 series, which was officially deprecated on June 30 to push users toward the newer generation. Built around a trillion-scale sparse architecture with roughly 1T total parameters and 42B active at inference, the model is engineered to handle extremely long horizons, including autonomous agent tasks that involve more than a thousand tool calls. Open weights are published under the XiaomiMiMo organization on Hugging Face, reflecting Xiaomi's commitment to letting developers self-host and fine-tune the model rather than relying solely on a closed API.

The flagship variant targets demanding agentic and software-engineering workloads, posting top placements on benchmarks such as ClawEval, GDPVal, and SWE-bench Pro, where Xiaomi positions it as rivalling top frontier agent systems on complex, multi-day professional tasks. Its 1M-token context window makes it a natural fit for integration with agent frameworks and coding assistants that need to reason over large codebases, long documents, or extended tool traces. Community experimentation has shown the model runnable across a pair of DGX Spark nodes, underscoring that the sparse activation pattern keeps the effective footprint manageable despite the trillion-parameter total scale, which broadens practical deployment options for teams with limited hardware budgets.