Hugging Face Models Now Available on Microsoft Foundry Managed Compute

Microsoft and Hugging Face jointly announced on July 7 that the Hugging Face Collection has gone live in the Microsoft Foundry Model Catalog, allowing enterprise users to deploy and run thousands of open-weight models on Foundry Managed Compute.

The core idea behind this collection is bridging the gap between the open-source model ecosystem and enterprise-grade deployment. Hugging Face hosts the world's largest open-source model community with models spanning text, image, audio, video, and more, while Microsoft Foundry provides enterprise infrastructure — identity management, network isolation, observability, security scanning, and automatic patch updates.

The deployment process is streamlined into five steps: browse the catalog and pick a model, choose a deployment template (latency- vs throughput-optimized, accelerator family, context length, quantization), configure instance count, deploy, and score via the unified Foundry endpoint with the familiar OpenAI SDK. The Azure AI team's Manoj Bableshwar and Osi detailed the workflow in the official blog post.

Deployment templates are the key design element of this release. Each template is a named, versioned asset that pins the runtime, accelerator family and count, context length, and model-specific tuning parameters. For Qwen3-32B, four templates are available side by side: single A100 with 40K context, single H100 with 40K context, dual A100 with 128K context, and dual H100 with 128K context — eliminating the need for manual runtime configuration.

Supported runtimes include vLLM (the primary GPU path), llama.cpp (for CPU and small-GPU GGUF deployments), TensorRT-LLM and NIM (for NVIDIA hardware with optimized kernels and Triton serving), and Hugging Face's own hf-serve (for vision, audio, segmentation, and other non-LLM pipelines covering every modality).

The service is available now in preview, supporting NVIDIA A100, H100, and AMD MI300X accelerators in both Global and Data Zone scopes. Each deployment comes with Playground support, Azure Monitor metrics, per-deployment billing tags, and automatic runtime upgrades with CVE patching.

This marks a significant infrastructure play from Microsoft: by bringing the Hugging Face open ecosystem into the Foundry enterprise platform, the company is promoting a new paradigm of "open models plus managed infrastructure." For enterprise users, this means flexible access to the latest open models while maintaining data sovereignty and compliance requirements.

The roadmap includes broader Hugging Face ecosystem coverage, additional accelerator families, and a Bring Your Own Weights (BYOW) feature for fine-tuned and proprietary variants deployed through the same templates and governance. Interested users can sign up for the preview at forms.cloud.microsoft/r/8Jnx1LALLA.