Skip to main content

Select, deploy, and evaluate Microsoft Foundry models

Foundry acts as a proprietary Huggingface, where as a developer you can:

  • browse through different types of models to find the ones that match your workflow
  • compare different models you've selected
  • deploy your model of choice so your applications can use it
  • evaluate the model with both "automated metrics and manual testing"

Explore the model catalog

Foundry's model cataloguecatalog includes two broad categories of models:

  1. Models sold directly by Microsoft through Azure
  2. Models provided for purchase through partners and for free through the community as OSS

Finding models in the catalog

Foundry exposes an "online shopping"-esque experience for its model catalog. Each model has a model card showing its key attributes such as the provider, its capabilities, any benchmark metrics for the model, any considerations to be made for responsible model usage, and deployment options.

  • Collection
    • Groupings of models that share a particular attribute, such as being provided by Huggingface vs Azure
  • Capabilities
    • The abilities of the model, such as reasoning (problem-solving capability), tool calling (the APIs and integrated functions the model can access), or multimodal processing (its ability to process text, images, audio, etc.)
  • Source
    • Who provides the model (i.e., OpenAI, Microsoft, Mistral, Anthropic, etc.)
  • Inference tasks
    • Specific abilities the model is tuned for, such as text generation, summarization, translation, image generation, speech synthesis, etc.
  • Fine-tuning methods
    • The ways developers can tweak and tune the model to their needs
  • Industry
    • Whether or not the model was trained on an industry-specific dataset

Understand generative AI model types

The catalog's offerings can be further subdivided into two additional groups:

  1. Large Language Models (LLMs) that are designed for tasks requiring deep reasoning, complex content generation, and extensive content understanding. They excel at sophisticated tasks but require more resources.
  2. Small Language Models (SLMs) that are designed for handling common tasks delivered via natural language. These models are efficient, cost effective, and can run on lower-end hardware, but are meant for simpler use cases.

Within these groups, there are:

Chat completion and reasoning models

Most models available in Foundry's catalog are chat completion models that are meant to generate coherent and contextually appropriate text responses, where they power conversational interfaces and content generation applications.

Within this family of chat completion models are reasoning models that provide enhanced problem-solving capabilities for fields like math, coding, strategy, and logistics. An example model that fit within this family is Claude Opus 4.6.

Specialized models

The Foundry catalog also includes models made with specific tasks in mind:

  • Embedding
    • Conversion of text into numerical representations, enabling semantic search. Language models actually leverage a form of this!
  • Image generation
    • Creation of images based on text descriptions
  • Video generation
    • Creation of video based on text descriptions
  • Image analysis
    • Generation of natural language output based on provided images and/or text
  • Text to speech
    • Conversion of textual input into synthesized speech
  • Speech to text
    • Transcribes audio into text
Regional and domain-specific models

Some models are optimized for and trained on specific languages or global regions. Think translation between languages.

Other models are trained specifically on medical literature, legal documents, or other industry materials.

Select models using benchmarks

Access model benchmarks

Back to top