Hugging Face Models on Foundry Managed Compute

Microsoft has announced the integration of Hugging Face models into its Foundry Managed Compute platform. This initiative provides a curated catalog of open-weight models from the Hugging Face ecosystem, updated weekly, and deployable with a single click onto Foundry Managed Compute. This brings enterprise-grade security, governance, observability, and billing to these open-source models, aligning them with other models available on Foundry.

Foundry Managed Compute acts as a managed GPU platform-as-a-service specifically for open-source and custom models. It automates machine maintenance, including container updates, runtime upgrades, and security patches, without requiring model redeployment. This ensures consistency across various deployment options, including pay-per-token and provisioned throughput, allowing developers to focus on model behavior rather than infrastructure.

Hugging Face is a central hub for open AI development, hosting millions of models and a vast community of builders. Open-source models have increasingly matched proprietary models in performance and offer unique advantages like customizability and offline operation. However, operationalizing these models for enterprise use has been a challenge due to complexities in discovery, licensing, security, and deployment.

The new Hugging Face Models on Foundry solution addresses these operational hurdles. A curated subset of Hugging Face models is brought into the Foundry Model Catalog through a systematic curation pipeline developed in collaboration with Microsoft. This ensures weights are pre-staged in Azure and runtimes are pre-built and scanned, allowing secure deployment within private networks and automatic upgrades of existing model deployments.