Hugging Face and Microsoft Foundry: Open Models on Managed GPUs

Hugging Face and Microsoft Foundry have launched a collaboration to host open-weight models from the Hugging Face ecosystem on the Foundry platform, which provides managed GPU compute as a service. The Hugging Face Collection is a curated subset of models refreshed weekly, covering text, vision, audio, and multimodal modalities, all security-screened and using the SafeTensors format. Models are deployed via templates that specify runtime, accelerator family, and context length, with runtimes including vLLM, SGLang, TensorRT-LLM, NIM, TEI, and llama.cpp. The curation pipeline involves identifying trending models, compliance and security screening, building and scanning runtime images, uploading weights to Azure storage, and validating API conformance and performance before publishing to the Foundry Model Catalog. Deployments are managed by Microsoft, with automatic container updates, runtime upgrades, and security patches. The service supports global and data zone deployments, and allows mixing open and frontier models within the same agent. The partnership aims to bridge the gap between open model flexibility and enterprise operational needs.