Deployment options
Foundry provides two deployment options:- Standard deployment in Foundry resources — For Foundry Models, including Foundry Models sold by Azure (also known as Azure Direct Models, or ADM) and select Models from partners and community. This option is the preferred and most capable deployment path.
- Managed compute deployment (preview) — Available for all Open Source Software (OSS) models, including models from partner and community, and custom models.
Standard deployment in Foundry resources
Standard deployment in Foundry resources is the preferred deployment option in Foundry. It supports the widest range of capabilities and deployment types.Which models use standard deployment?
All Foundry Models, including Foundry Models sold by Azure and select Models from partners and community use standard deployment. Foundry Models sold by Azure include all Azure OpenAI models and selected models from top providers that are billed through your Azure subscription, covered by Azure service-level agreements, and supported by Microsoft. Select Models from partners and community that use standard deployment include Anthropic models, and specific models from partners like Mistral, Cohere, and Meta.Capabilities
Standard deployment supports:- Multiple deployment types — Global Standard, Data Zone Standard, Regional Standard, Provisioned, Batch, and more. Each type controls where data is processed and how you pay. For details, see Deployment types for Microsoft Foundry Models.
- Data processing flexibility — Choose regional, data zone (US or EU), or global processing based on your compliance requirements.
- Content filtering — Built-in Azure AI Content Safety filters with customizable configurations.
- Keyless authentication — Microsoft Entra ID (recommended) and key-based authentication.
- Private networking — Virtual network integration for secure access.
- Provisioned throughput — Reserve capacity with PTUs for predictable, low-latency performance. For details, see Provisioned throughput.
Resource requirements
Standard deployment is available in:- Foundry resources — The primary resource type for new Foundry projects. No AI Hub required.
- Azure OpenAI resources — If you use Azure OpenAI resources, the model catalog shows only Azure OpenAI models for deployment. Upgrade to a Foundry resource for access to the full set of Foundry Models.
Managed compute deployment (preview)
Managed compute in Foundry is currently in public preview and registration is required to use it.
This preview is provided without a service-level agreement, and we don’t recommend it for production workloads. Certain features might not be supported or might have constrained capabilities. For more information, see Supplemental Terms of Use for Microsoft Azure Previews.
Managed compute supports open-source, partner, industry, and custom models. Managed compute deployments are served on the unified Foundry project endpoint, using the same authentication, networking, and SDK surface.
Which models use managed compute?
Examples of model collections that require managed compute include:- Hugging Face
- Some Meta models
- Some Mistral models
- NVIDIA inference microservices (NIMs)
- Industry models (Saifr, Rockwell, Bayer, Cerence, Sight Machine, Page AI, SDAIA)
- Databricks
- Custom models
Capabilities
Managed compute (Preview) supports:- Unified Foundry endpoint and authentication — Use the same project endpoint, API keys, Microsoft Entra ID, and private networking as pay-per-token and provisioned throughput deployments. Inference routes use
<endpoint>/managed-deployments/<deployment-name>/. Chat-completions-compatible runtimes also work on the standard/openai/v1/route with the OpenAI SDK. - Model-instance sizing — Deployments are sized in model-centric terms. You don’t need to pick virtual machine SKUs, because Foundry chooses GPUs per instance based on model size, architecture, context length, and whether the workload is optimized for latency or throughput.
- Optimized inference runtimes — Microsoft-curated vLLM, SGLang, and NVIDIA NIM containers with continuous batching, speculative decoding, tensor parallelism, and LoRA hot-swap.
- Accelerator families — A100 (80 GB), H100 (80 GB), H200 (141 GB), and MI300X.
- Auto-scaling and scale-to-zero — Auto-scale from live traffic or scale manually. Configure an idle timeout so the deployment scales to zero when no traffic arrives, making billing stop immediately.
- Microsoft-managed runtimes — Microsoft owns serving runtimes, base container images, and security patches. Updates are applied to live deployments automatically.
- Observability metrics — Each deployment emits API call count by status code and response-time percentiles. Chat-completion models also emit input and output token counts, time-to-first-token (TTFT) percentiles, and total response-time percentiles, grouped by time.