Microsoft and Hugging Face expand collaboration
Hugging Face
Microsoft and Hugging Face just made 10,000+ open models one-click deployable on Azure AI Foundry. That's up from a much smaller catalog two years ago, and more models (video, 3D, agents) are coming fast.
Based on reporting by Hugging Face — 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
Microsoft used its Build conference stage to widen a partnership that started two years ago as a modest bridge between Azure and Hugging Face's model hub. Back then the hub held roughly 200,000 open models. Today it's closing in on 2 million, and Azure AI Foundry's Model Catalog now surfaces more than 10,000 of them under a dedicated Hugging Face Collection, ready to spin up with a couple of clicks.
The pitch is straightforward: developers pick whatever model fits the task — text generation, image segmentation, speech recognition, translation, you name it — and deploy it on Azure's infrastructure without leaving the enterprise security perimeter. That last part is the real hook. Every model in the Collection has to clear Hugging Face's own vetting, including scans from ProtectAI Guardian and JFrog, and ship in safetensors format rather than the older Pickle format, which has a history of letting malicious code hitch a ride. Microsoft and Hugging Face say they'll keep re-scanning inference containers going forward, not just at launch.
The workflow itself is almost boring in a good way. Take Microsoft's own Phi-4 Reasoning Plus model as an example: find it in the Collection, hit Deploy, choose a VM size and instance count, and you're running. You can also start from the Hugging Face model page directly and jump into Azure Machine Learning Studio through a "Deploy on Azure ML" button. No ceremony, no separate account gymnastics.
What's more interesting than today's catalog size is the roadmap. Hugging Face says day-zero releases from major model providers will land in Azure Foundry the same day they hit the hub, and trending models will get pulled in on a rolling, near-daily basis rather than in batch updates. New modalities are on the list too — video, 3D, time series, even protein models — alongside a push toward smaller, specialized models built specifically for agent workflows. Asha Sharma from Microsoft framed it as giving developers freedom of choice while keeping enterprises in control; Hugging Face's Clement Delangue put it more bluntly, calling it a way for companies to "take control of their AI destiny" instead of renting someone else's black box.
My take — AI-written commentary, not fact-checked reporting
This is Microsoft doing what it does best: embracing open source just enough to make Azure the default place people run it, without actually open-sourcing anything of its own. Good for developers who want optionality, fine for enterprises who want a vetted supply chain, but let's not pretend this is a victory for decentralization — it's a very effective moat dressed up as generosity.
Read more about this at: Hugging Face