Xiaomi opens the weights of the MiMo-V2.6 series

The weights are out in the open. Xiaomi's MiMo team has released the MiMo-V2.6 series on Hugging Face as open weights (a model anyone can download and run on their own machine). It is MIT-licensed and free.

Pro-RL, Flash-RL and Distill-Qwen-9B arrived on 21 September (22 September, China time). The MOPD versions were added about a week later. Pro is a 1.02-trillion-parameter MoE; only 42 billion of those parameters do any work on a given token. Flash has 309 billion in total, with 15 billion active. Both have a context of 1 million tokens. Besides text, they also read images, video and audio, and they give their answers as text. The MOPDs are versions distilled from several expert teacher models, with fewer repeated tool calls. The 9B is built on Alibaba's Qwen3.5-9B. Xiaomi presents it as a starting point for research, not as a ready-made assistant.

A note for anyone planning to run these at home: in MXFP4 format, Pro takes up about 566 GB and Flash 173 GB. No independent party has verified Xiaomi's own benchmark results yet (89.9 on Terminal Bench 2.1, for example). On the Artificial Analysis index, Pro leads the open models with 46 points, but it trails closed models such as GPT-6 Sol (48). On the same index its cost per task is about one eighth of GPT-6 Sol's, and its speed is less than half.

The download is free. Running it isn't.