The phone company just became one of the most interesting AI labs on the planet.
On September 22, 2026, Xiaomi's AI team released MiMo-V2.6-Pro: a 1.02-trillion-parameter AI model that anyone can download, fine-tune, and use commercially — completely free, under the MIT license. And it debuted as the highest-scoring open-weight AI model in the world.
What MiMo V2.6 actually is
Pro is a sparse mixture-of-experts model: 1.02 trillion total parameters, but only about 42 billion "active" per token — so it's far cheaper to run than its headline size suggests. It's natively omnimodal, meaning the same model handles text, images, video, and audio instead of bolting vision on afterward, and every checkpoint ships with a 1-million-token context window.
The family comes in four flavors: Pro (1.02T total / 42B active) for the hardest agentic work, Flash (309B / 15B active) for the best balance of brains and cost, Pro UltraSpeed for when you need Pro-level answers fast, and a 9B distilled model that runs on a single GPU — the one tinkerers will actually play with.
The number that matters: $2.6 million
Xiaomi says the Pro model's reinforcement-learning run took under six days and cost roughly $2.62 million, with the Flash variant at about $850,000. Not long ago, frontier AI training runs were discussed in the hundreds of millions. The team is led by Luo Fuli, a former DeepSeek researcher — talent is flowing, and it isn't all flowing to Silicon Valley.
Xiaomi also open-sourced the training code and more than 7,000 reinforcement-learning environments, so researchers can reproduce the results themselves. That kind of transparency is rare at this level.
What it scored
Pro hit 46 on the Artificial Analysis Intelligence Index — the highest score of any open-weight model (it ties Grok 4.7 and beats DeepSeek V4.1 Pro's 36), and within touching distance of closed frontier models. Pricing is the real punch: $0.435 per million input tokens and $0.87 per million output tokens, with a measured cost per benchmark task roughly one-twentieth to one-sixtieth of leading international models. Xiaomi kept pricing unchanged from V2.5 — the score went up and the bill didn't.
Why it matters
Open-weight models this strong, this cheap, and this permissive reshape the economics of building with AI. A startup can now fine-tune a frontier-class model without paying OpenAI or Anthropic rates. MiMo-V2.6 won't dethrone Claude or GPT overnight — it still trails the best closed models — but it's the DeepSeek-R1 moment of 2026: it resets what "expensive" means in AI. And the open training stack is the real gift.
Comments
Post a Comment