A trillion-parameter model under an MIT licence
Xiaomi released the MiMo-V2.6 series on Monday and put the weights on Hugging Face under an MIT licence. The flagship, MiMo-V2.6-Pro, is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion active per token. A cheaper MiMo-V2.6-Flash has 309 billion total and 15 billion active. Both take text, images, audio and video, and both carry a one-million-token context window.
Artificial Analysis, which runs its own evaluations rather than republishing vendor tables, scored MiMo-V2.6-Pro at 46 on its Intelligence Index and ranks it first of the 114 open-weight models it tracks. The median open-weight model on that index sits at 18.
The price is the argument
On Xiaomi’s API the Pro model costs $0.43 per million input tokens and $0.87 per million output, with a 99 per cent discount on cached input. Flash costs $0.14 and $0.28. Artificial Analysis puts the cost of pushing its whole Intelligence Index through Pro at about $0.13 a task, which places the model on its intelligence-against-cost frontier.
That is roughly a fifth of what the nearest Western comparison charges. SpaceXAI’s Grok 4.7, which Artificial Analysis scored at the same 46 when both landed on Monday, is priced at $2 per million input tokens and $6 per million output.
The independent index also records the practical trade-offs. Pro generates at 125 tokens a second, twelfth fastest of the open-weight models tracked, and takes 2.19 seconds to its first token. It is verbose: the evaluator needed 140 million tokens to run the full index through it.
Trained where anyone could watch
Xiaomi framed the release as built in public. Since 15 September a live dashboard has streamed the reinforcement learning runs for both models — steps, reward curves, token throughput, benchmark scores and a running cost counter.
By Xiaomi’s own accounting the two models each completed 30 reinforcement learning steps across roughly 750,000 trajectories, at about $2.62m for Pro and about $850,000 for Flash. Xiaomi says it used two techniques to stop the models gaming their own reward — one that builds a rubric per task by comparing several attempts at it, another that shifts training weight towards the better passing solutions inside a rollout group. Those are the company’s own descriptions of its own method; no outside party has reproduced them.
What is open and what is not
The MIT licence covers the weights, and Xiaomi is also publishing the technical report, the reinforcement learning training environments and the RL code. That is a wider release than Alibaba managed last week, when Qwen-Image-2.1 arrived with open weights but a research-only licence.
A third variant, MiMo-V2.6-Pro-UltraSpeed, is a hosted service rather than a download: Xiaomi sells it at a premium for up to twenty times faster output.
What to watch
The number to watch is whether the ranking survives contact with evaluators other than Artificial Analysis. The second is deployment: an MIT licence on a one-trillion-parameter model is only useful to organisations with the hardware to serve 42 billion active parameters, and most of the demand will land on the hosted API and on routers rather than on anyone’s own GPUs.
