Azure AI Foundry Model Catalog: How to Choose Among 10,000+ Models

More than 10,000 models and 50 new ones every month -- Azure AI Foundry's catalog has passed the point where browsing it manually makes sense. What it needs is a method, not just scrolling.

The real scale of the catalog in 2026

The Microsoft Foundry Models catalog now includes more than 10,000 models, with roughly 50 new ones published every month -- growth that makes any "review the whole catalog" strategy a non-starter. The catalog features models from providers like Azure OpenAI, Anthropic, Mistral, Meta, Cohere, NVIDIA, and Hugging Face, including models trained by Microsoft itself.

How to actually filter

You can search and discover models that fit your needs through keyword search and specific filters: by provider, task, industry, and capability (reasoning, tool calling, and more). That's the right starting point -- never browse the full catalog without filters applied from the outset.

The model categories available

Models range from general-purpose foundation models, reasoning models, small language models (SLMs), and multimodal models, to domain- and industry-specific models -- the sheer variety of categories is itself a signal that "the best model" depends entirely on the specific task you're solving, not on some universal ranking.

The built-in leaderboard as a shortcut

The model catalog also offers a model performance leaderboard and benchmark metrics for selected models, accessible via "View leaderboard" and "Compare models" -- instead of digging through scattered external benchmarks, Foundry gives you a direct comparison inside the same interface where you'll deploy the model.

A practical selection strategy

With a catalog this size, the strategy that works is: filter first by required capability (reasoning, tool calling, multimodal), then by provider if you already have an ecosystem or compliance preference, and only then use the leaderboard to pick among the final 3-5 candidates -- trying to compare 10,000 options from scratch is a guaranteed recipe for analysis paralysis.