Xiaomi releases MiMo-V2.6-Pro-RL open-weight multimodal model

Xiaomi MiMo released downloadable weights for MiMo-V2.6-Pro-RL, along with a model card and linked technical report. The repository labels the release with an MIT license.

Xiaomi says the model processes text, images, video and audio, with a maximum context length of 1 million tokens. It describes MiMo-V2.6-Pro-RL as a sparse mixture-of-experts model containing 1.02 trillion parameters in total and activating 42 billion for a given input.

According to the model card, the system has 384 routed experts and activates eight at a time. Its 70-layer language-model backbone comprises 60 sliding-window-attention layers and 10 global-attention layers.

The disclosed architecture also includes a 681-million-parameter MiMo vision transformer, a 308-million-parameter AudioTokenizer encoder and a 127-million-parameter audio patch encoder. Xiaomi describes a five-layer speculative decoder that predicts seven subsequent tokens per forward pass for parallel verification.

The model card’s benchmark tables contain Xiaomi’s own evaluation results, not independently verified rankings. The evidence reviewed for this story included no independent reproduction of those scores.

Xiaomi says the model is also available through Xiaomi AI Studio, MiMo Code, Xiaomi MiMo Desktop, the Xiaomi MiMo Open Platform API and OpenRouter. The research for this story did not independently test those services.

One specification remains unresolved: the model card gives MiMo-V2.6-Pro-RL a total parameter count of 1.02 trillion, while the Hugging Face interface separately displays a model size of 524 billion parameters. The available sources do not explain the difference.