The Model CatalogThe searchable library inside Azure AI Foundry that lists all available AI models from Azure, partners, and community, with filter controls for task type, provider, deployment option, and capabilities. is the central hub for discovering and deploying AI models within the Foundry portal. It contains over 1,900 models from providers including Azure OpenAI, Meta, Mistral, Cohere, NVIDIA, Hugging Face, DeepSeek, and Microsoft itself.
The catalog is organized into two top-level categories:
- Azure Direct Model (ADM)A model hosted, billed, and supported by Microsoft under Microsoft Product Terms. ADMs support Standard (serverless) deployment and do not require an Azure Marketplace subscription. (Azure Direct Models / ADMs) — fully hosted, billed, and supported by Microsoft under Microsoft Product Terms
- Models from Partners and Community — governed by the model provider's own license terms and billed through Azure Marketplace (subscription required)
Each model has a Model CardThe detail page for a single model in the catalog. Contains Quick facts (context window, version, provider), Details, Benchmarks, Deployments, License, and Artifacts tabs. showing quick facts, benchmarks, and deployment options. When you deploy a model you assign a Deployment NameThe label you assign when deploying a model. Your application code uses this name — not the model name — in the model parameter of API calls. The deployment name can differ from the model name. that your application code uses in API calls.
Model versions are identified by a date-stamped string (e.g., 2024-11-20 for GPT-4o). When a new default version is released, deployments with auto-upgrade policies can switch automatically; deployments set to NoAutoUpgrade continue on the pinned version until it is retired.
Catalog Filters — When to Use Each
To narrow down candidates you use the catalog's filter panel:
| Filter | Use when… |
|---|---|
| Inference tasks | You know the task type (chat, embeddings, image-to-text, etc.) |
| Collection | You need a specific provider (Azure OpenAI, Meta Llama, Mistral, etc.) |
| Deployment options | You must use serverless (pay-per-token) or managed compute (VM-based) |
| Capabilities | You need reasoning, tool-calling, or multimodal support |
| Industry | You need a model pre-trained on medical, legal, or domain-specific data |
| Fine-tune tasks | You plan to fine-tune the model |
| Licenses | You have license compliance requirements |
Combining an Inference tasks filter with a Collection filter is the fastest way to find the right model for an exam scenario.
Model Card Tabs
| Tab | What it shows |
|---|---|
| Quick facts | Context window, release date, provider, latest version |
| Details | Full description, version history, supported data types |
| Benchmarks | Standardized performance metrics (MMLU, HumanEval, etc.) for that model only |
| Deployments | Your existing deployments of this model |
| License | Provider license terms |
| Artifacts | Asset download (open-weight models only) |
Use the catalog's Compare models (or View leaderboard) feature for side-by-side comparisons — individual model cards only show that model's own benchmarks.
Deployment Options by Model Family
| Deployment Option | Billing model | Available for |
|---|---|---|
| Standard (serverless) | Pay-per-token | ADMs and select partner models |
| Managed compute | VM core-hours | Open-weight (Hugging Face, NVIDIA NIMs), custom models |
| Provisioned throughput (PTU) | Reserved capacity per hour | Select ADMs with high-volume needs |
| Global Batch | Pay-per-token, async | Select ADMs for offline batch processing |
Partner/community models require subscribing your project to the Azure Marketplace offering and accepting license terms before the first deployment.
Open the catalog — Azure AI Foundry portal (ai.azure.com) → Discover (upper-right navigation) → Models (left pane). The full catalog is displayed with a search bar and filter panel on the left.
Filter by task — In the filter panel, expand Inference tasks → check Chat completion (or your target task). The catalog grid updates in real time.
Narrow by provider — Expand Collection → check Azure OpenAI (or Meta, Mistral, etc.) to see only that provider's models.
Filter by deployment type — Expand Deployment options → check Standard to show only serverless-capable models, or Managed compute for open-weight/custom models.
Open a model card — Click a model tile (e.g., gpt-4o) → review the Quick facts panel (context window, latest version, license). Switch to the Benchmarks tab to compare MMLU or HumanEval scores.
Compare models — From the catalog home, select Compare models → pick 2–4 candidates → view side-by-side benchmark results to make a data-driven selection.
Check version details — On the model card, select the Details tab → scroll to Version info to see available versions, deprecation timelines, and supported data types.
AI-3016 Assessment Focus
Catalog questions test your ability to choose the correct filter for a task, identify which deployment option applies to a given model family, and understand partner model subscription requirements.
Exam Trap
"Any model in the catalog can be deployed as serverless." Only selected ADMs and specific partner models support Standard (serverless/pay-per-token) API endpoints. Hugging Face open-weights and NVIDIA NIMs require managed compute (dedicated VM).
Exam Trap
"The deployment name is the same as the model name." The deployment name is set by you at deploy time. Your application always references the deployment name in the model parameter — not the model name itself.
Exam Trap
"Models from Partners and Community require no additional subscription." These models are offered through Azure Marketplace. You must subscribe your project to the model offering and accept license terms before the first deployment.
Exam Trap
"You can compare models side-by-side using the Benchmarks tab on any model card." Individual model cards show only that model's own benchmarks. Use the catalog's Compare models feature for side-by-side analysis.
Exam Tip
The fastest filter path for a task-specific search: Inference tasks first (narrows by what the model does), then Collection (narrows by who provides it).
Must Memorize
Hugging Face models and NVIDIA NIMs = Managed compute only (no serverless option). Azure OpenAI models = Standard (serverless) or Provisioned (PTU).
Question — click to flip
Q: A developer needs text embeddings for semantic search. Which catalog filter narrows results most directly?
Question — click to flip
Q: Which deployment option is required for open-weight Hugging Face models in Azure AI Foundry?
Question — click to flip
Q: What additional step is required before deploying a Cohere model from the Azure AI Foundry catalog for the first time?
Question — click to flip
Q: In Azure AI Foundry, what identifier must your application code use when calling a deployed model?
Question — click to flip
Q: Which tab on a model card shows standardized performance scores like MMLU and HumanEval?
Question — click to flip
Q: What is the top-level organizational split in the Azure AI Foundry model catalog?