Nvidia signs the Hugging Face deal at $12.9 billion
The rumors are over. Nvidia confirmed Thursday that it has a definitive agreement to acquire Hugging Face, the closest thing the open-model world has to a town square — and on the same day, a UAE research institute shipped what it calls the largest fully open model fleet in history, with training data attached.
Nvidia has agreed to acquire Hugging Face for $12,930,300,000, ending a week of speculation and putting the chipmaker in ownership of the platform where most of the world's open models live. Jensen Huang announced the deal in a blog post Thursday, saying Hugging Face "will remain an open platform for the entire AI ecosystem" and that "NVIDIA compute will not be required to build on or deploy through Hugging Face." Clément Delangue framed the sale as a scale decision rather than a surrender: open models need "more compute, more support, more collaboration, and more visibility. That's why we went to talk to Jensen."
The numbers explain why Nvidia wanted this. Huang says more than 18 million developers use Hugging Face to share over 3 million models, 500,000 datasets and 1 million applications, and more than 200,000 companies use it to discover and deploy AI. That is distribution, not just a code repository — and it lands days after G20 nations endorsed a light-touch AI framework, with Nvidia making a very public case that open weights are a national-security asset. Huang made that argument explicitly, saying cybersecurity firms "couldn't do it without open models" and that open models are "vital to the American economy."
Our read: the promise of openness is the whole asset here, which is exactly why it should be treated as a commitment to audit, not a reassurance. Hugging Face's value to Nvidia is that developers trust it as neutral ground — the moment it visibly favors Nvidia silicon or Nvidia models, that trust is the thing that gets sold for $12.9 billion and then quietly spent. Watch whether the multi-accelerator language becomes structural (board independence, a published neutrality policy) or stays a blog paragraph. We argued the case for skepticism when this was still just talks — Deep Dive — Nvidia's quiet takeover of the AI stack.
MBZUAI's Institute of Foundation Models shipped K2 Horizon, six models from 0.9B to 375B-A23B, with weights, code, training data and training recipes all released. IFM calls it the largest fully open-source model launch in AI history, and the claim that matters isn't the parameter count — it's the packaging. Where most "open" releases ship weights and a license, IFM is publishing the data and the methodology too, plus intermediate checkpoints so researchers can watch capability emerge across training rather than only at the end.
The fleet spans genuinely different hardware targets: a 0.9B model aimed at watches and glasses, 3.7B and 7B for phones, a dense 32B and sparse 36B-A4B for on-prem servers, and a 375B-A23B for enterprise. Two technical claims stand out — a "diffusion distillation" technique that generates blocks of tokens in parallel for roughly 3× faster generation, and a mixture-of-value attention design that improves reasoning without extra compute. The 36B-A4B also carries a native 512K context.
The local-AI community's reaction has been cautiously enthusiastic, and the caution is warranted: these are the vendor's own benchmarks, and nobody outside IFM has independently reproduced them yet. But the transparency is the interesting part. Artificial Analysis found the big model attempts only 40% of its omnicience eval questions, declining the rest rather than guessing — a low hallucination rate earned by abstention, not by knowing more. Shipping the training data is what makes that claim checkable, and that's a higher bar than the open-weights debate has been setting lately.
Google DeepMind released WeatherNext 3, which forecasts at 5km resolution every hour instead of every six — and improves rain prediction 60% over its predecessor. The model beats deep-learning rivals from Google, Microsoft, Nvidia and the European Center for Medium-Range Weather Forecasting, as well as the traditional physics-based forecasts from the US National Weather Service and the ECMWF. It's already feeding weather results in Google Search, Maps and Gemini, with cloud access for researchers.
Three specific fixes got it there. The model is 2.4 times larger than WeatherNext 2, it targets individual weather stations rather than averaged grid cells — which makes its output checkable against ground truth — and it ingests raw satellite observations hourly instead of waiting for supercomputer-processed analysis. The practical upshot is mundane in the best way: weather is one of the few AI applications that touches nearly every person's day, and hourly, neighborhood-level rain forecasts are the difference between a useful warning and a shrug.
What to watch: whether Nvidia's neutrality pledge for Hugging Face gets written into anything binding, and whether independent testers can reproduce IFM's K2 Horizon numbers now that the training data is public.
Should a chipmaker be allowed to own the neutral hub where open models are distributed — even with a promise to keep it neutral? Tell us in the comments.
