MiMo V2.6 Pro is the flagship of Xiaomi's open-source MiMo-V2.6 series, a multimodal family engineered around agentic and professional workloads. Xiaomi positions the series as a step in its recursive self-improvement research, scaling reinforcement learning compute on verifiable complex tasks during a six-day Live RL training cycle that built on roughly half a year of prior research. The Pro variant is described as a sparse mixture-of-experts model with around 1.02 trillion total parameters, paired with a lighter Flash sibling at roughly 309B parameters, and the series as a whole is marketed for "flagship performance, full modality, built for professional workflows."
In third-party reporting on the Artificial Analysis Intelligence Index, MiMo-V2.6-Pro is cited at a 46-point composite score, placing it ahead of comparable open-source systems like Kimi K3 and Qwen3.8 Max, while still trailing leading closed-source systems such as Claude Fable 5.1 and GPT-6 Astra. Both Pro and Flash support multimodal input and a one-million-token context window, making the family well suited to long-horizon planning, coding assistance, and tool-heavy agent loops. The weights are released under an MIT license, so the model can be self-hosted or routed through compatible inference gateways that pass upstream cache discounts through to the caller.
