Microsoft launches Foundry Managed Compute: thousands of Hugging Face models in one click

On July 7, 2026, Microsoft launched Foundry Managed Compute, a managed GPU platform for deploying open Hugging Face models on the Azure cloud. Thousands of models from a curated catalog now deploy in one click, with enterprise security, monitoring and unified billing.

What is Foundry Managed Compute

Foundry Managed Compute is a platform service built on managed GPUs. It picks the GPU topology and runs the infrastructure for you. Developers think in terms of the model: parameter count, context length and the trade-off between latency and throughput. Microsoft handles the servers.

Microsoft pitches Foundry itself as an agentic AI stack with the widest model selection of any cloud provider. It offers models from OpenAI, Anthropic, Meta, Mistral, DeepSeek and Hugging Face through a single endpoint and SDKs for Python, C#, JavaScript and Java. The platform includes multi-agent orchestration, end-to-end tracing, content safety filters, an AI Red Teaming Agent and Azure Policy integration.

Which Hugging Face models are now available?

The Hugging Face collection in Foundry covers every key modality:

• Catalog updates: weekly, driven by community trends and customer requests
• Formats: SafeTensors only, with no executable third-party code
• Runtimes: vLLM, SGLang, TensorRT-LLM, NIM, Text Embeddings Inference, llama.cpp, hf-serve
• Accelerators in preview: NVIDIA A100, H100 (80 GB), AMD MI300X
• Modalities: LLMs, multimodal models (VLMs), ASR, embeddings, segmentation, image generation

Every model goes through a five-stage selection pipeline: trend detection → license and repository security checks → runtime build with CVE scanning → weight upload to Azure Storage → API validation and publication to the catalog.

Microsoft stores prepared weights and runtime images in Azure Storage ahead of time, so production deployments don't need outbound access to the Hugging Face Hub. Security patches and runtime updates apply automatically, without redeploying the model.

How to deploy a model

Deployment takes five steps: pick a model in the catalog, choose a deployment template, set the instance count, launch through the portal, CLI, SDK or REST, and call the model through the unified Foundry endpoint. Qwen3-32B, for example, comes with four templates: vLLM paired with either A100 or H100 accelerators and either a 40K or 128K token context.

With the Python SDK, launching takes a single `begin_create_or_update` call. You then call the model through the standard OpenAI-compatible SDK, exactly as you would a frontier model from OpenAI or Anthropic. Foundry agents integrate Hugging Face models through the same paths as closed models.

"We've combined the breadth of the open-source ecosystem with an operational layer that

Microsoft runs" — from the partners' official announcement.

What it means

The integration lowers the barrier to enterprise use of open models. Instead of configuring GPU clusters, runtimes and security systems themselves, companies get a managed service with unified Azure RBAC access policies, Azure Monitor monitoring and per-deployment billing. Hugging Face brings together 15 million developers, 400,000 organizations and more than 3 million open models, and that entire ecosystem now becomes an enterprise asset on Azure infrastructure.

FAQ

Do you need access to the

Hugging Face Hub during deployment?

No. Microsoft preloads model weights and runtime images into Azure Storage, so deployment itself makes no outbound requests to the Hugging Face Hub. That matters for enterprise environments with restricted internet access.

How often is the model catalog updated?

The catalog grows weekly, based on signals from the Hugging Face community and direct customer requests. Every new model passes through the five-stage security pipeline before publication.

*Meta is designated an extremist organization and banned in the Russian Federation.