Xiaomi opens MiMo-V2.6 under MIT: the 1.02T-parameter Pro leads open weights
In one sentence On September 22 Xiaomi published on Hugging Face, under MIT, MiMo-V2.6-Pro (1.02 trillion-parameter MoE, 42B active) and MiMo-V2.6-Flash (309B, 15B active), plus a 9B distill, more than 7,000 reinforcement learning environments and training code. Pro scores 46 on the Artificial Analysis index, a record for an open model.
Xiaomi, best known for smartphones, released the MiMo-V2.6 model family with downloadable weights under the MIT license, which allows commercial use. The largest model, MiMo-V2.6-Pro, accepts text, images, video and audio and has a one-million-token context.
According to Artificial Analysis, Pro scored 46 on its intelligence index, the highest score ever reached by an open-weight model; its predecessor stopped at 26. The smaller Flash costs just 14 cents per million input tokens via API.
Beyond the models, Xiaomi published more than 7,000 reinforcement learning environments and the code to reproduce training: rare material that lets other labs repeat the process. The release came days after Anthropic named Xiaomi among seven Chinese labs accused of distilling Claude's outputs; Xiaomi did not comment.
Companies
Xiaomi
Tools
MiMo-V2.6-Pro, MiMo-V2.6-Flash, MiMo-V2.6-Distill-Qwen-9B
Tags
Sources
- https://thenextweb.com/news/xiaomi-mimo-v2-6-open-weight-model-anthropic-distillation
- https://www.eweek.com/news/xiaomi-mimo-v26-open-source-rl-reproduction/
- https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader
- https://technode.com/2026/09/22/xiaomi-open-sources-mimo-v2-6-models-after-scaling-reinforcement-learning/