At Microsoft Build 2026, Microsoft introduced Foundry Managed Compute alongside a curated catalog of open-weight models from Hugging Face on Microsoft Foundry. The new offering provides a managed GPU platform-as-a-service designed to bring enterprise security, governance, observability, and billing to open-source models. According to the announcement, weights are pre-staged in Azure, and runtimes are built, scanned, and automatically updated by Microsoft without requiring redeployments. Foundry Managed Compute serves as the third deployment option in Foundry, sitting alongside pay-per-token and provisioned throughput. Users can deploy models by specifying workload characteristics such as parameter count, context length, and optimization goals, while Foundry manages the underlying GPU topology. The platform integrates a versatile collection of community-built inference runtimes selected and tuned for specific model architectures, including vLLM, SGLang, Text Embeddings Inference, llama.cpp, TensorRT-LLM, NIM, and hf-serve. Because weights and runtime images live in Microsoft-managed registries within Azure, deployments can occur inside private networks without outbound network access to the Hugging Face Hub. Models deployed from the Hugging Face Collection slot directly into Foundry Agents and can be accessed through unified endpoints using the OpenAI SDK. The Hugging Face Collection in the Microsoft Foundry Model Catalog is available now in preview.
Hugging Face Models on Foundry Managed Compute
At Microsoft Build 2026, Microsoft announced Foundry Managed Compute and a curated catalog of Hugging Face open-weight models on Microsoft Foundry. The platform allows users to deploy Hugging Face models in one click with enterprise security, governance, and billing. The integration provides a managed operational layer for open-source models using runtimes like vLLM, SGLang, and TensorRT-LLM.
Why it matters
This integration bridges the gap between open-source AI development and enterprise production requirements by providing managed infrastructure, security, and unified endpoints for Hugging Face models. It removes traditional operational hurdles such as GPU sizing, CVE patching, and runtime selection.