Hugging Face and Microsoft are tightening an already productive partnership with two announcements aimed squarely at developers who want open models without infrastructure headaches. The first is an expansion of the Hugging Face Collection inside the Azure Model Catalog, which now includes Meta’s Llama 3, Mistral AI’s Mistral 7B, Cohere’s Command R Plus, Qwen 1.5 110B, and several top-performing community fine-tunes from the Open LLM Leaderboard. The pitch is simple: find a model on the Hub, hit “Deploy on Azure,” and it spins up in your own Azure account. Or skip the Hub entirely and browse the collection directly in Azure AI Studio.
Since launching a year ago, the Hugging Face Collection has drawn hundreds of Azure AI customers and grown to over a thousand open models. That’s real traction, but the missing piece was always the marquee names — the models developers actually ask for by name. Adding Llama 3 and Mistral 7B changes the calculus. These aren’t niche experiments; they’re the default starting points for a huge share of open-source AI work right now.
The second announcement pairs Hugging Face’s optimization work with Microsoft’s new Azure ND MI300X virtual machines, built on AMD Instinct MI300 GPUs and now generally available. Hugging Face says it collaborated with AMD and Microsoft on performance tuning through its Optimum-AMD library, including ROCm integrations and continuous testing. Specific benchmark numbers aren’t disclosed in the announcement, but the implication is that open models run notably faster and cheaper on this hardware than they would out of the box.
There’s also a developer-experience play worth watching. Spaces Dev Mode, now in beta for PRO subscribers, lets users connect a Hugging Face Space to local VS Code or a web-hosted VS Code environment. Given that the community has already built over 500,000 demo apps on Spaces, lowering the friction between prototyping and real development could push more of those experiments toward production. Microsoft, meanwhile, continues to lean into open models on its own terms — nearly 300 models in its Hugging Face organization, including the MIT-licensed Phi-3 family. The 3.8-billion-parameter Phi-3 mini outperforming many 7-10B models is a legitimate flex, and Hugging Face has already put it into Hugging Chat. The throughline across all of this: Microsoft wants Azure to be where open models actually get deployed, not just downloaded and forgotten.
