Microsoft Foundry adds curated Hugging Face open-weight models with managed GPU compute

Microsoft and Hugging Face announced that a curated subset of open-weight models from the Hugging Face ecosystem is now deployable in one click on Microsoft Foundry Managed Compute, a managed GPU platform-as-a-service. The new Hugging Face Collection is refreshed weekly and includes models across text, vision, audio, and multimodal categories such as LLMs, VLMs, ASR, speech translation, embeddings, segmentation, and image generation.

Every model in the Collection is security-screened, ships in SafeTensors format, and avoids untrusted code execution paths unless rigorously reviewed. Microsoft builds and scans inference runtimes (vLLM, SGLang, TensorRT-LLM, NIM, TEI, llama.cpp) for CVEs, signs the images, and publishes them to a Microsoft-managed registry. Weights are pre-staged in Azure, and each model is paired with a runtime optimized for its workload.

From the developer’s perspective, an open-weight model in the Hugging Face Collection behaves identically to other models in the Foundry Model Catalog: it uses the same endpoint, SDKs (Python, C#, JavaScript, Java), authentication, observability, and billing. Microsoft Foundry supports pay-per-token, provisioned throughput, and Managed Compute as three deployment options, all unified under a single bill and API surface.

The offering includes enterprise controls such as content safety filters, task-adherence guardrails, an AI Red Teaming Agent for adversarial testing, unified RBAC, private networking, and Azure Policy integration. Foundry also provides end-to-end tracing, real-time monitoring, continuous evaluations, and a prompt optimizer that improves agent behavior based on evaluation results.

The curation process identifies trending models from community signals, partner requests, and customer demand, then screens for license compliance and security before building and scanning runtimes. Microsoft Foundry is positioned as a platform for building and operating agentic AI applications, with a model catalog spanning Microsoft, OpenAI, Anthropic, Meta, Mistral, DeepSeek, Hugging Face, and others, and a Foundry Agent Service that adds multi-agent orchestration with built-in memory, knowledge grounding via Foundry IQ, and a catalog of connectable tools.

Managed Compute lets developers deploy a model instance described by parameter count, context length, and optimization target (latency or throughput), while Foundry handles the underlying GPU topology and automatic updates to containers, runtimes, and security patches without redeploying the model.