The most interesting AI lab in the world right now might be a phone company. Xiaomi's MiMo team has released MiMo-V2.6 — two open-weight models, Pro and Flash — and the flagship MiMo-V2.6-Pro has just debuted at the top of Artificial Analysis's Intelligence Index with a score of 46, ahead of every other open model on the planet, including DeepSeek's best.
MiMo-V2.6-Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters (42 billion active per token). The smaller Flash variant carries 309 billion total parameters with 15 billion active. Both are natively multimodal — text, image, video, and audio — with a one-million-token context window.
The headline grabber is the price tag. Xiaomi says the reinforcement-learning run behind Pro took under six days and cost about $2.62 million. Flash cost roughly $850,000. Not long ago, frontier training runs were talked about in hundreds of millions of dollars; Xiaomi did this one for the price of a nice house. And under an MIT license, anyone can download the weights from Hugging Face, fine-tune them, self-host them, even resell them — no revenue cap, no research-only clause.
API pricing is just as aggressive: $0.435 per million input tokens and $0.87 per million output tokens for Pro. Artificial Analysis measures the cost per task at roughly one-twentieth to one-sixtieth of leading international models. The team is led by Luo Fuli, formerly of DeepSeek — the open-weight talent pipeline is flowing fast.
Xiaomi went beyond the checkpoint drop: it published around 7,000 reinforcement-learning training environments and the end-to-end training framework, so researchers can study and reproduce the recipe. The models are live on Hugging Face, Xiaomi's API platform, AI Studio, MiMo Code, MiMo Desktop, and OpenRouter.
Why it matters: an MIT-licensed model at the top of the open leaderboard changes the economics of building with AI. Startups can now self-host a frontier-class model instead of renting tokens by the million, and every closed provider has to justify its premium on reliability or features. The Chinese open-weight race — Kimi, GLM, Qwen — just got its leaderboard reset.
One caveat: independent testing (Artificial Analysis) supports Pro's competitiveness, but Xiaomi's self-reported coding benchmark scores deserve healthy skepticism until third parties replicate them in their own harnesses.
The takeaway: open weights are no longer 'almost as good.' At the top end they're a genuine alternative, with a cost story closed labs can't match. The era of the nine-figure training run as a moat is quietly ending.
Comments
Post a Comment