Explore the model catalog
Unit 2 of 7 · Module 2: Select, deploy, and evaluate Microsoft Foundry models
Go deeper
Unlocks when you've watched the lesson.
Read this lesson
Two piers
Think of the Foundry Models catalog as a huge showroom, the place for discovering models and comparing them side by side. With more than ten thousand on display, the real skill is narrowing it down. Everything in it falls into two groups. Models sold directly by Azure are billed right through your Azure subscription, like buying from the store itself. That group covers the Azure OpenAI models, plus models from Microsoft and other providers. The other group comes from partners and the community, and each of those has its own licensing and pricing, more like a marketplace seller. Whichever you look at, its model card is the label on the box: who made it, what it can do, how it scores on benchmarks, responsible AI notes, and how you can deploy it.
Search and filter
In the Foundry portal, you can type a keyword to search, then narrow things down with filters. Collection: say, just the models provided directly in Azure, or just the ones from Hugging Face. Capabilities: reasoning through hard problems, tool calling so the model can use your APIs and functions, or multimodal input, like text, images, and audio together. Source, which is the provider: Azure OpenAI, Microsoft, Mistral, Cohere, Meta, Anthropic, and others. Inference task, meaning the job you want done, such as generating text, summarizing, translating, making images, or synthesizing speech. Fine-tuning: which ways a model can be customized with your own data. And industry: models trained on data from one field, which often outperform general-purpose models inside that field.
Large, small, and thinking
Browse a little, and you'll notice language models come in different sizes and styles. Large language models, such as GPT-5, Mistral Large, or Llama 3 70B, are the heavyweights: deep reasoning, long and complex writing, and room for lots of context. They excel at demanding work, but they need more computing power, and cost more to run. Small language models, such as Phi-4, Mistral's open models, or Llama 3 8B, are the nimble ones, efficient and affordable for everyday language tasks. Pick one when speed and cost matter more than top-tier reasoning; some are compact enough for lower-end hardware or edge devices, like a phone or a factory sensor. The kind you'll see most is the chat completion model: you send it a conversation, it writes a natural, fitting reply, and that's what sits behind chat apps and content writing. And for tough problems in math, code, science, or planning, reasoning models like Claude Opus 4.6 slow down, break the problem into steps, and can show their reasoning.
The specialists
Beyond the talkers, the catalog has specialists, each built for one kind of job. Embedding models, like OpenAI's text embedding three, in large and small sizes, or Cohere's embedding models, turn text into lists of numbers that capture its meaning. That's what powers semantic search, recommendations, and retrieval augmented generation, or RAG: ask about a refund policy, and it can find the line that says returns are accepted within thirty days, with no words in common. Image generation models, like GPT image two, draw pictures from a text description, handy for ads, illustrations, or design mockups. Video generation models, like Sora, do the same for video. Image analysis models, like GPT five mini, look at a picture and your question together, and answer in natural language. Text-to-speech models, like GPT four o mini TTS, read text aloud in a synthetic voice. Speech-to-text models, like GPT four o mini transcribe, go the other way, writing down what was said. Finally, some models specialize in particular languages, regions, or industries, trained on sources like medical research or legal documents. When your app lives in that niche, they often beat general-purpose models.
Facts from Microsoft's documentation: Microsoft Foundry Models overview (CC BY 4.0), rewritten in Loop's own words and examples. Narration, animation, and quiz by Loop. Microsoft, Azure, and Microsoft Foundry are trademarks of Microsoft. Loop is independent and not affiliated with or endorsed by Microsoft.