Microsoft Foundry now offers a curated Hugging Face Collection—thousands of open-weight models across text, vision, audio, and multimodal modalities, refreshed weekly, deployable onto Foundry Managed Compute.

Microsoft pre-stages weights in Azure, builds and CVE-scans runtimes (vLLM, SGLang, TensorRT-LLM, NIM, TEI, llama.

cpp, hf-serve), and validates each model+accelerator combination for API conformance and performance before publishing it to the Foundry Model Catalog with a one-click deploy path.

Models are security-screened (SafeTensors only, no untrusted code unless rigorously reviewed), and their licenses are reviewed against Microsoft’s enterprise distribution policy.

Deployments don’t require outbound network access to Hugging Face Hub—weights are served from Azure storage.

The developer surface is unified: a single endpoint, the same SDKs (Python, C#, JavaScript, Java), the same authentication, the same observability, and a single bill shared with pay-per-token and provisioned throughput models.

Developers pick a deployment template that pins the runtime, accelerator family/count, context length, and tuning—no GPU topology decisions.

The Collection integrates with Foundry Agents the same way frontier models do. Available now in preview on NVIDIA A100, H100, and AMD MI300X accelerators, in Global and Data Zone scopes.

On the roadmap: broader ecosystem coverage, additional accelerator families, and Bring Your Own Weights for fine-tuned variants.