AI NewsModels & agentsAnnouncement
Xiaomi released MiMo-V2.6-Pro-RL, a 1 trillion parameter mixture-of-experts model, under an MIT licence
Xiaomi published MiMo-V2.6-Pro-RL to Hugging Face on 22 September 2026, a 1.02 trillion parameter sparse mixture-of-experts model with 42 billion active parameters, a 1 million token context and an MIT licence.
Image: Hugging Face, via XiaomiMiMo
Why it mattersAn MIT licensed open-weights model at this size, with the whole reinforcement learning stage described in one run, gives a team building a coding or agent stack a full-precision weight file it can inspect, fine-tune and self-host.
A 1.02 trillion parameter model whose weights anyone can download and fine-tune, and whose licence file says MIT, went up on Hugging Face this week. On 22 September 2026 the Xiaomi MiMo team published MiMo-V2.6-Pro-RL, a sparse mixture-of-experts model with 42 billion active parameters per token, together with a MiMo-V2.6-Flash-RL variant at 159 billion and a MiMo-V2.6-Distill-Qwen-9B distillation.
The shape of the model card
The Hugging Face model card describes the LLM backbone as 70 layers of a mixed sliding-window and global-attention design, with 384 routed experts and 8 activated per token, and a 1 million token context length. A MiMo ViT vision encoder of 681 million parameters, plus a 308 million parameter audio tokeniser and a 127 million parameter audio patch encoder, make the model omnimodal on text, images, video and audio in a single set of weights. A five layer speculative decoder ships with it.
Xiaomi's own benchmark claims
The model card reports Xiaomi's own scores on nine benchmarks. On code-agent work, MiMo-V2.6-Pro scores 71.9 on DeepSWE v1.1 and 63.2 on Xiaomi's own MiMo Code Bench. On general agent work, 76.9 on Toolathlon-Verified, 82.0 on OSWorld-Verified and 62.0 on JobBench. Two cybersecurity numbers stand out, at 94.0 on CyberGym and 80.2 on the MiMo Cyber Bench. The model card labels these as Xiaomi's own results, and it does not report figures from an independent evaluator.
What Xiaomi says about the training run
Two claims from the model card are worth reading in the original. Xiaomi describes what it calls "Scaling RL Compute: Fully asynchronous Group Relative Policy Optimization (GRPO) on very large batches, 1,568 prompts times 16 rollouts per step, billions of tokens per update." Xiaomi also says the coding, general-agent, visual and cybersecurity training was combined into a single mixed run, which the card names the "You Only RL Once" recipe. The training data, token count and compute bill are not disclosed on the model card.
The independent numbers to watch for
Xiaomi has not published a paper or a peer-reviewed evaluation to go with the release. The Latent Space AINews issue that first flagged the release, on 22 September 2026, cites a $2.6 million figure for the reinforcement learning phase from one analyst, and reports an Artificial Analysis intelligence-index score of 46, both of them second-hand.
For a team choosing a hosted or self-hosted open-weights model, the story is a full-precision 1 trillion parameter checkpoint under an MIT licence, with a technical report PDF in the repository and no clause forbidding commercial use. Whether the benchmark numbers survive independent testing is the next thing to watch.
Source
- XiaomiMiMo/MiMo-V2.6-Pro-RL, model card on Hugging Face
- MiMo-V2.6 technical report PDF
- Latent Space AINews, 22 September 2026, for the analyst cost figure and Artificial Analysis score
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
