[AINews] Xiaomi MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M
Latent Space ● Covered by 4 sources
Xiaomi just dropped MiMo-V2.6-Pro, a new open-weights model it says is its most capable yet. It’s native omnimodal, and the RL stack behind it is unusually open.
Based on reporting by Latent Space — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Xiaomi has joined the frontier-model race with something that sounds a lot less like a phone launch than it should. MiMo-V2.6-Pro is the headline model in a new series, and Xiaomi says it is its most capable model so far. There’s also MiMo-V2.6-Flash for a better speed-cost balance, plus MiMo-V2.6-Pro-UltraSpeed, which the company says can produce output up to 20x faster at the same quality for people who need extreme generation speed.
What makes this release stand out is not just the model name. Xiaomi is presenting MiMo-V2.6 as natively omnimodal, and in a field where a lot of “open” releases still feel carefully boxed in, that matters. The bigger surprise is the company itself: Xiaomi is not usually counted among the familiar Chinese AI leaders, which makes this a rather awkward entrance for the old club.
The technical report leans hard into reinforcement learning. Xiaomi says it scaled RL compute along three axes: larger batches and higher throughput, more tasks and richer environments, and more grader compute. The setup included fully asynchronous training, 1,568 samples per update, context lengths of up to 1M, and 3.5 to 3.7B tokens per step. The task mix spans coding, general agents, visual and cyber work, with the stated goal of making gains in one area help the others.
And Xiaomi is not keeping the machinery to itself. It says the tooling, including the environments, will be open sourced, along with the environment code and training recipes. The complete 7k-plus task datasets have not yet been released, but the company is already promising code, loaders, rewards, scorers, adapters, and other bits of the stack. That’s the real tell here: the model is the headline, but the RL infrastructure is the strategic move.
The release also arrives with an unusual amount of transparency. A former DeepSeek engineer now at Xiaomi had been publishing final RL training runs live, and the source material says that helped expose internal metrics in a way people are not used to seeing. Artificial Analysis says MiMo-V2.6-Pro starts at the top of its Intelligence Index among open weights models, which is a neat way of saying Xiaomi did not just ship a research demo and hope nobody looked too closely.
My take — AI-written commentary, not fact-checked reporting
This is the kind of open release the industry actually needs: not just weights for bragging rights, but environments, recipes, and enough detail to copy the method. The uncomfortable part for closed-model cheerleaders is that the moat keeps shrinking where it matters most — at the point where systems get trained, tested, and improved. Fancy secrecy is a weaker sales pitch when someone else is openly publishing the plumbing.
Read more about this at: Latent Space
