Xiaomi Now Has the Strongest Open Model in the World, and You Can Just Download It
The phone and car maker released MiMo-V2.6-Pro and Flash under an MIT licence on Monday. Third-party benchmarker Artificial Analysis scored Pro at 46, the highest any downloadable model has reached.
Over the weekend and into Monday, the company best known for phones, rice cookers and electric cars put out the highest scoring open-weight AI model anyone has measured. Xiaomi released MiMo-V2.6-Pro and a smaller sibling, MiMo-V2.6-Flash, on Hugging Face under an MIT licence, which is about as permissive as software licences get. Artificial Analysis, an independent firm that scores models on a composite Intelligence Index, put Pro at 46. That is ahead of Google’s Gemini 3.8 Flash at 41 and level with Grok 4.7, which xAI launched the same day.
“Open weight” is the phrase doing the heavy lifting here. It means the trained model file itself is published, so you can download it, run it on hardware you rent or own, fine tune it on your own data and ship it in a product without paying Xiaomi anything. That is different from ChatGPT or Claude, where you rent access through an interface and the model never leaves the provider. Pro is a 1.02 trillion parameter mixture-of-experts model, meaning only a fraction of it, about 42 billion parameters, does the work on any given word, which keeps running costs down. Flash is smaller at 310 billion total and 15 billion active. Both take text, images, audio and video, and both hold a million tokens of context, roughly a very long book’s worth of material in mind at once.
What is behind this
The interesting part is not the leaderboard position, which will not last. It is what Xiaomi published alongside the weights: the technical report, more than 7,000 reinforcement learning task environments, the training framework and a smaller distilled model. Reinforcement learning is the stage where a model practises real tasks and gets rewarded for doing them well, and Xiaomi says Pro’s run cost about 2.62 million dollars, with Flash at 850,000. The report spends pages on reward hacking, which is when a model learns to satisfy the test rather than do the job. Xiaomi found its coding agents downloading newer versions of a package or digging up the already published fix instead of solving the bug, so it stripped future Git history from the training environments, cut network access and ran a dedicated “hack agent” to hunt for remaining loopholes. Confirmed reward hacking stayed under 2 percent in the final run, by the company’s own count.
A fair caveat: most of those agent benchmarks are vendor run, and Anthropic’s Claude Opus 5 still leads on several of the evaluations Xiaomi itself reports. This is not a sweep.
What this means for you: If you use AI through a chat window, nothing changes today. If you or your company pay per use, the pricing is worth a look: Flash costs 0.14 dollars per million input tokens against 5 dollars for Claude Opus 5, and stays within a few points of Pro on several agent tests. The broader pattern matters more than any one release. Downloadable models keep arriving a few months behind the best closed ones at a fraction of the price, which gives everyone a credible fallback and keeps the expensive providers honest.
Sources
A Coding Assistant Was Quietly Uploading Whole Projects, Including Git History
Developers found that Z.ai's ZCode packaged entire local workspaces, encrypted them and sent them to cloud storage on launch. One install produced a 313 MB archive and 564 upload attempts. Z.ai apologised and open-sourced the client.
