Azure AI Foundry

Foundry is also designed to support an AI solution throughout its entire lifecycle. The overall system is built around, among other things, the following capabilities:

• Model availability: Foundry offers users thousands of language and other models from both Microsoft and Hugging Face. Developers can choose which ones to deploy. Models like GPT‑5, DeepSeek, Mistral and thousands of others are equally accessible.
• Model comparison: Within Foundry you’ll find extensive benchmarking data that lets you investigate how different models perform in various scenarios. This helps developers pick the most appropriate model for each case. The newest, most advanced model might give the best results but could also be unnecessarily expensive and slow. With model comparison you can find the simplest model that delivers adequate quality.
  
• Model testing: Foundry supports running models in a test bench, for example by feeding specific inputs through different models and examining how well or appropriately they respond. The same testing tools can be used to examine content‑security boundaries, for example by testing how well a model adheres to given behavioural rules. Testing also matters in the ongoing lifecycle of the project: for instance when upgrading model versions it’s wise to perform careful testing so that a seemingly benign version upgrade doesn’t compromise result quality.
• Model deployment: Large general‑purpose models like GPT‑5 are always run in the cloud as shared resources. Smaller language models, for example Microsoft’s Phi, can also be run in your own copy. Foundry offers the option to deploy chosen models in dedicated environments. The basic idea is: whatever model you need, you can get it from Foundry at the click of a button whether as a shared service or even as a dedicated virtual machine.
• Model fine‑tuning: With Foundry you can also create fine‑tuned versions of models. Sometimes it is difficult to get a language model to behave exactly as you wish because its base training “pushes” it strongly. Fine‑tuning allows you to feed the model examples of how you want it to behave. This isn’t necessary in most projects, but when the need for control is significant and success is critical, fine‑tuning helps. In Azure AI Foundry you can perform fine‑tuning directly via the interface and of course also deploy the fine‑tuned models for your applications.
• AI governance support: Foundry includes features specifically designed to manage model usage. For example, you can test model behaviour against defined rules, and export the results to external AI‑governance systems. 
  Microsoft Learn
• Security tools: AI Foundry also supports, for instance, content moderation (via Azure Content Safety) and attack‑defence measures. With techniques like “Prompt Shield”, Foundry can identify when someone attempts to override base instructions of the model. These safeguards can be flexibly configured for your application.