Skip to main content
With a basic understanding of generative AI and AI agents, this lesson shows how to get started with generative AI in Azure. It focuses on Microsoft Foundry models: how to discover available foundation models and what to consider when choosing and deploying them. Azure provides a range of capabilities to help you select AI models, connect them to applications, and deploy them in a controlled, secure way. This guide walks through the model catalog in Microsoft Foundry, explains model origins, and outlines the operational factors you should evaluate before deployment.

​
Discovering models in the Microsoft Foundry model catalog

Think of the Model Catalog like an online library — instead of books, it contains AI foundation models. You can browse available models, compare them, and select the one that best fits your application’s requirements. The catalog exposes a large collection of models with varied capabilities such as chat completion, code generation, text classification, and other inference tasks.
A screenshot of the Microsoft/Azure Foundry Models catalog showing a searchable list of AI models and filter options. The interface displays model names (e.g., claude-opus, gpt-5.2, DeepSeek) and buttons like "View leaderboard" and "Compare models."
The catalog includes filtering options to narrow results by capability, model source, inference task, fine-tuning methods, and industry tags. Use the comparison feature to evaluate key differences (accuracy, latency, instruction-following, cost, and domain alignment) rather than choosing solely on model popularity. For example, a customer support application will likely prioritize:
  • strong natural language understanding,
  • instruction-following accuracy,
  • predictable and safe outputs,
  • cost/performance that fits your SLA.
Use the catalog filters and comparisons to match a model’s strengths to these priorities.

​
Where models originate

Microsoft Foundry aggregates models from multiple sources. Knowing a model’s origin matters for billing, support, security, and compliance.
A slide titled "Microsoft Foundry Models" showing two columns: "Microsoft-Managed Models" and "Partner and Community Models." The left lists Microsoft-offered models billed through Azure (including Azure OpenAI) and the right notes models from trusted third-party providers offered via a model catalog.
Key source categories:
  • Microsoft-managed models: Hosted and operated by Microsoft, billed through Azure subscriptions (this may include Azure OpenAI and other Microsoft-provided foundation models).
  • Partner and community models: Trusted third-party providers listed in the Foundry catalog; these widen choice and enable comparisons across vendors.
Table — model source considerations: Different models specialize in different tasks (free-form conversation, reasoning, code generation). The catalog helps you compare those trade-offs in one place.

​
Deployment and operational considerations

Choosing a model is only the first step. Before deploying it into production, evaluate these operational factors to ensure your solution meets both functional and non-functional requirements.
A slide titled "Microsoft Foundry Models" showing "Model Deployment Considerations" with four labeled boxes: Deployment Type, Model Version, Rate Limits, and Content Safety and Guardrails. Each box includes a short descriptor about where the model runs/pricing, version and lifecycle, tokens/requests per minute, and filtering/safety policies.
Important factors to evaluate:
  • Deployment type
    Where and how the model runs (cloud-hosted, managed inference, on-premises) affects latency, availability, data residency, and billing. Choose the option that aligns with your latency and compliance needs.
  • Model version
    Models evolve and different versions can change behavior, latency, and lifecycle rules. Pin to a specific version for predictable behavior and define an upgrade path for improvements or security fixes.
  • Rate limits and throughput
    Understand quotas such as tokens per minute or requests per minute. Plan capacity with autoscaling, batching, retry/backoff strategies, and caching where appropriate.
  • Content safety and guardrails
    Implement filtering, moderation, and governance controls to mitigate harmful or inappropriate outputs. Integrate enterprise-wide security and compliance requirements into your deployment.
Table — deployment trade-offs:
When evaluating models, balance capability with operational requirements: deployment location, version stability, throughput limits, cost, and safety controls. These considerations help you build systems that are reliable, scalable, and responsible.

​
Summary

Microsoft Foundry’s model catalog helps you discover and compare foundation models from Microsoft and trusted third parties. When selecting and deploying a model, be deliberate about:
  • which model best fits your application needs,
  • where the model is hosted and billed,
  • which model version to pin to,
  • throughput and rate-limit planning,
  • and the safety and governance controls required.
Next steps: pick a candidate model from the Foundry catalog, validate it against representative workloads, and plan deployment with monitoring, scaling, and safety controls.

Watch Video