Hugging Face announced that a curated collection of open-weight models is now available in preview on Microsoft Foundry Managed Compute. The integration, introduced at Microsoft Build 2026, allows organizations to deploy pre-screened models onto Azure infrastructure through a single endpoint and standard SDKs in Python, C#, JavaScript, and Java.
Microsoft and Hugging Face refresh the catalog weekly, adding models across text, vision, audio, and multimodal tasks. Microsoft pre-stages all model weights in Azure storage and hosts runtime images in a Microsoft-managed container registry. Because the assets reside directly in Azure, deployments do not require outbound network access to the Hugging Face Hub and can run inside private enterprise networks.
Security and runtimes
Hugging Face reported that every model undergoes license and security screening before publication. Repositories must use the SafeTensors weight format, and Microsoft excludes or remediates any model that relies on unverified executable code or third-party Python scripts at load time. Microsoft builds, scans for vulnerabilities, and signs container images across several inference engines, including vLLM, SGLang, Text Embeddings Inference, llama.cpp, TensorRT-LLM, and NIM.
The platform applies runtime upgrades and security patches automatically without requiring users to redeploy existing model instances. Models integrate directly with the Foundry Agent Service, task-adherence guardrails, an AI Red Teaming Agent, and Azure Policy controls.
Hardware options
Foundry Managed Compute operates as a managed GPU platform-as-a-service alongside pay-per-token and provisioned throughput options. Organizations pay for accelerators by the hour and can scale capacity to zero when workloads are idle. Compute options include Nvidia A100, Nvidia H100, and AMD MI300X accelerators deployed across Global capacity scopes or Data Zone scopes for data residency.
Deployment templates allow operators to select pre-tuned configurations for context length, accelerator count, and latency or throughput targets. Hugging Face stated that future releases will expand model coverage, introduce additional accelerator families, and add a Bring Your Own Weights feature for custom fine-tuned variants.
