At Microsoft Build 2026, the company announced Foundry Managed Compute and a new Hugging Face Collection on Foundry. The collection is a curated catalog of open-weight models from the Hugging Face ecosystem. Models are refreshed every week and can be deployed in a single click onto Foundry Managed Compute. Weights are pre-staged in Azure, runtimes are built and scanned by Microsoft, and every model in the collection comes with the same enterprise security, governance, observability, and billing that applies to other models on Foundry.
The Platform: Microsoft Foundry and Managed Compute
Microsoft Foundry is a platform for building and operating agentic AI applications. It offers the widest model selection on any cloud, including models from Microsoft, OpenAI, Anthropic, Meta, Mistral, DeepSeek, and Hugging Face. All models are accessible through a single endpoint and a single set of SDKs in Python, C#, JavaScript, and Java.
On top of those models sits the Foundry Agent Service, which provides multi-agent orchestration with built-in memory, knowledge grounding through Foundry IQ, and a catalog of connectable tools via agentic protocols. Agents can work with enterprise data. Once agents are running, Foundry provides end-to-end tracing, real-time monitoring, continuous evaluations, and a prompt optimizer that improves agent behavior based on evaluation results.
Developers also get content safety filters, task-adherence guardrails, an AI Red Teaming Agent for adversarial testing, unified RBAC, private networking, and Azure Policy integration directly within the platform.
Alongside pay-per-token and provisioned throughput, Foundry Managed Compute is the third deployment option. It is a managed GPU platform-as-a-service for open-source and custom models. Users deploy a model instance described by parameter count, context length, and optimization for latency or throughput. Foundry handles the GPU topology underneath, so users think and plan in model terms.
Microsoft takes care of the machine: container updates, runtime upgrades, and security patches happen automatically on supported runtimes including vLLM, SGLang, TensorRT-LLM, NIM, TEI, and llama.cpp. Users do not need to redeploy their models. Model configuration, deployment behavior, and routing stay with the user.
Pay-per-token, provisioned throughput, and Managed Compute share a single endpoint, the same SDKs, the same authentication, the same observability, and a single bill. Open-source models integrate with Foundry Agents the same way frontier models do, so users can mix model types in a single agent without a separate integration path.
Managed Compute offers global deployments for broad capacity and best pricing, as well as Data Zone deployments for residency and sovereignty. The same code and workflow apply. Quota is aligned to accelerator families, so a plan built on the H100 family today carries forward as new hardware generations come online.
Why Hugging Face
Hugging Face is the public square of open AI. It has 15 million builders, 400,000 organizations, and over 3 million open models published. New frontier capabilities such as agentic coding, video segmentation, speech, and embeddings land weekly. It is the GitHub of open models, where the community publishes weights, writes model cards, compares evaluations, and pulls models for experimentation.
Open models have closed the gap with proprietary models on benchmark after benchmark. They unlock capabilities that proprietary endpoints cannot. State-of-the-art is now open. Leading open-weight models are competitive with top closed frontier models on widely used benchmarks. Full weights make deep customization possible: fine-tuning, distillation, quantization, and LoRA adaptation. Weights run in the user's tenant on infrastructure they control. Users can shape costs by paying for accelerators by the hour, scaling to zero when idle, and right-sizing GPUs to the specific model. Version control allows pinning a specific model version, evaluating it, deploying it, and moving forward or rolling back on the user's own release cadence.
The catch has always been the operational layer: discovery, license review, security screening, runtime selection, GPU sizing, image building, CVE patching, and standing the model up behind an enterprise-grade endpoint. Hugging Face by itself is not an enterprise serving platform. Hugging Face models on Foundry is that operational layer, run by Microsoft.
Hugging Face Models on Foundry
The Hugging Face Collection brings a curated subset of models directly into the Foundry Model Catalog. Models are refreshed weekly. Trending models from the Hugging Face ecosystem are added continuously as the community publishes them. The collection covers every modality: text, vision, audio, and multimodal. It includes LLMs and VLMs for chat and agents, ASR and speech translation, embeddings, segmentation, and image generation.
All models use the SafeTensors weight format with no untrusted code. Every model in the collection is security-screened. There are no trust_remote_code execution paths unless rigorously reviewed. The right runtime is matched to the model: vLLM and SGLang for LLMs, TensorRT-LLM and NIM where applicable, TEI for embeddings, llama.cpp for CPU. From the user's side, an open-weight model in the Hugging Face Collection looks and behaves like any other model in the Foundry Model Catalog. Every model in the collection has been put through a multi-stage publishing pipeline before appearing in the catalog.
The Curation Pipeline
Hugging Face and Microsoft work together to bring popular open-weight models to Microsoft Foundry. They identify trending models based on community signals, partner requests, and customer demand. Candidates are selected for enterprise readiness.
Models are screened for compliance and security. Licenses are reviewed against Microsoft's enterprise distribution policy, and license metadata is captured and preserved on the catalog model card. Repositories are inspected for trust_remote_code patterns and custom executable code. Any model that would require executing third-party Python at load time is either remediated or excluded.
Microsoft builds inference container images on supported runtimes, including vLLM, SGLang, TensorRT-LLM, NIM, TEI, and llama.cpp. These images are scanned for CVEs, signed, and published to a Microsoft-managed container registry. Model weights are pulled from Hugging Face once, validated against the published model card, and stored in Microsoft-managed Azure storage in the regions where the model is served.
Stay updated
Get the day's AI and automation news in your inbox. No spam, unsubscribe anytime.
Every model, runtime, and accelerator combination is tested for API conformance and performance. The validated model, with its templates, runtime images, and weights, is published to the Foundry Model Catalog with a one-click deploy path onto Managed Compute. Because weights are pre-staged in Azure storage and runtime images live in a Microsoft-managed registry, deployments do not need outbound network access to Hugging Face Hub. Users can deploy to production inside a private network.
Model Runtimes
Hugging Face models on Foundry are powered by a collection of community-built, open-source inference runtimes. Each runtime is selected and tuned for Foundry Managed Compute and matched to the model architectures it serves best. Across all runtimes, new versions and patches land on Foundry quickly, and existing model deployments are upgraded automatically without requiring the user to redeploy.
vLLM is the default high-throughput serving engine for open large language models, tuned for production GPU workloads. Hugging Face is a direct contributor to vLLM, so any model in the Transformers library can run on vLLM out of the box. When a new model lands on Hugging Face, it can be served on Foundry the same day with no waiting on a custom integration.
SGLang is a serving engine for language and multi-modal models. It has strong support for structured outputs such as JSON, regex, and grammar-constrained generation, which agentic and tool-using workloads depend on. Hugging Face and the SGLang team have built a Transformers backend integration for SGLang, so any model in the Transformers library runs on SGLang out of the box and reaches Foundry the same day it lands on Hugging Face.
Text Embeddings Inference (TEI) is the runtime for embedding, reranker, and sequence-classification models. Accelerator-specific images ship with kernels compiled for each GPU and CPU family Foundry supports, keeping the embedding hot path lean for RAG and semantic-search workloads.
llama.cpp is the CPU and small-GPU path for GGUF-quantized models. It is useful for cost-optimized deployments, smaller models, and CPU-only regions, with the same OpenAI-compatible API as vLLM and SGLang.
TensorRT-LLM and NIM are used on NVIDIA hardware where NVIDIA's optimized kernels and Triton-based serving deliver better latency or throughput for specific model families.
hf-serve is Hugging Face's own multi-model inference server, used for model architectures outside the LLM and embedding fast paths. It covers vision, audio, segmentation, and other Transformers-native pipelines, so the collection can cover every modality with a consistent serving layer.
Deploying and Scoring an Open-Weight Model
The Hugging Face Collection in the Foundry Model Catalog is where users start. Deployment takes five steps. First, browse the catalog and pick a model. The deploy wizard surfaces the model ID, deployment template ID, and acceleratorType needed for scripting via SDK or REST. Second, choose a deployment template: latency-optimized or throughput-optimized, accelerator family, context length, and quantization. Third, configure the instance count to scale throughput by adding model instances. Fourth, deploy from the portal, CLI, SDK, or REST. Fifth, score via the unified Foundry endpoint with the SDK already in use.
A deployment template is the unit of choice in step two. It is a named, versioned asset that pins the runtime, accelerator family and count, context length, and runtime-specific tuning. Picking a template is the only knob users turn for how they want the model to run.
For example, qwen3-32b ships with four templates: one on A100 with 40K context, one on H100 with 40K context, one on 2x A100 with 128K context, and one on 2x H100 with 128K context. Each template arrives pre-tuned for the model, with runtime settings, tool-call and reasoning parsers, scoring path, health probes, request concurrency, and any model-specific context-extension settings all set by Microsoft. Trade-offs are called out inline in the template description. When users script the deploy, they reference the template and Foundry handles the rest.
The deploy can be done via Python SDK using the Cognitive Services Management client. The deployment is reachable through the unified Foundry endpoint with the OpenAI SDK. The model field takes the deployment name. A chat-completions model from the collection slots into Foundry Agents as an admin-connected model and is callable through the Foundry Responses API with the same OpenAI SDK, same auth, same endpoint, and same observability.
What's Available Today
The Hugging Face Collection in the Microsoft Foundry Model Catalog is available now in preview. It includes thousands of models across every modality, refreshed weekly. Models can be deployed onto Foundry Managed Compute with NVIDIA A100, NVIDIA H100, or AMD MI300X accelerators in Global and Data Zone scopes. There is a unified Foundry endpoint with Playground support, first-class Azure Monitor metrics, per-deployment billing tags, and curated runtime upgrades and CVE patching applied automatically to deployments.
Microsoft is accepting sign-ups for the preview at forms.cloud.microsoft/r/8Jnx1LALLA. On the roadmap are broader coverage of the Hugging Face ecosystem, additional accelerator families, and Bring Your Own Weights for fine-tuned and proprietary variants deployed through the same templates and governance as collection models.
Hugging Face is where open models are published and discovered. Microsoft Foundry is where enterprises operationalize them on curated, license-screened, security-screened weights hosted in Azure. Runtimes are community-built and CVE-scanned. A single endpoint provides enterprise identity, networking, observability, and agent integration. The breadth of the open-source ecosystem is combined with the operational layer Microsoft runs underneath.
Related on Neura Market
- Explore the AI Tools Directory for more enterprise AI platforms and model deployment solutions.
- Check the Cloud Computing Marketplace for other managed GPU services and infrastructure offerings.
- Visit the Open Source AI Models Directory to discover other open-weight model ecosystems and hosting options.