MiMo-V2.6-Pro-RL

MiMo-V2.6-Pro-RL is the flagship open model of the MiMo-V2.6 series from Xiaomi, released on 2026-09-21. According to the model card, it has about 1.02 trillion parameters in total, of which 42 billion work on each step, and a context window of one million tokens. It is one model that takes text, images, video and audio.

The authors say it was trained with one large reinforcement-learning run across programming, general agent tasks, vision and cyber security, instead of separate runs for each area, and then polished by learning from several stronger teacher models. The goal is long, multi-step work for an AI agent, such as large code repositories and long tool-use traces. A smaller sibling, MiMo-V2.6-Flash-RL, is released next to it and is not covered by this card.

The weights are public without any acceptance step. The license is MIT, which allows commercial use. We found no territorial restrictions in the license. The card gives launch examples for SGLang and vLLM; the vLLM example splits the model across 8 accelerators, and the SGLang example uses two servers. The model is meant for server-level hardware. The card compares it with other models in its own table; these are the developer's figures, not independent checks.

Date of check: 2026-10-21.