Xiaomi MiMo releases MiMo-V2.6-Pro-RL, a 1.02T omnimodal MoE under MIT

Xiaomi MiMo published MiMo-V2.6-Pro-RL, which the card calls the flagship checkpoint of the MiMo-V2.6 series. It is a sparse mixture-of-experts model with 1.02T total and 42B activated parameters, a 1M-token context, and input in text, image, video and audio. It includes a vision encoder and an audio encoder, and a 5-layer multi-token-prediction head for speculative decoding.

The card's emphasis is the training method. It describes one mixed reinforcement-learning run across coding, general agents, visual and cybersecurity tasks, fully asynchronous GRPO, and an agentic grader that compares rollouts within each group. A final multi-teacher on-policy distillation stage follows. These are the maker's descriptions; the card links a technical report.

The licence is MIT. Smaller siblings in the series are listed on the card, including a Flash variant.

The card gives no release date; the date used is the Hugging Face repo creation date.

More of the card is about how it was trained than what it is for. The model is still real, and the licence is the friendliest at this size.

Sources

• Model card on Hugging Face
• Hugging Face repo metadata (creation date, licence, size)

Editor's first pass (AI-assisted)

• Real 3.5: MIT licence and ungated weights; 1.02T parameters is a cluster-sized download.
• Useful 3.5: Coding and agent work are the target, with audio and video input as an extra.
• Notable 3.5: 764 likes, modest next to the biggest releases here; the card calls it the flagship of its series.
• Novel 3.5: The RL recipe (groupwise agentic grading, multi-teacher distillation) is the distinctive part.