AI-901
Deep Dive
Microsoft Foundry is Microsoft's central platform for building generative AI solutions. It unifies model access, evaluation tools, deployments and monitoring in a single portal. The AI-901 tests its capabilities and those of the integrated model catalog.
Contents
Microsoft Foundry (formerly Azure AI Studio, then Azure AI Foundry) is Microsoft's unified platform for designing, building, evaluating and deploying enterprise-grade generative AI applications. Accessible at ai.azure.com, with a toggle between the "New" and "Classic" experience.
Model catalog
Access to hundreds of models from Microsoft, OpenAI, Meta, Mistral, Cohere and more — all in one place.
Integrated playgrounds
Experiment with models without writing code. Compare responses, tune parameters, test system prompts.
Managed deployments
Publish models as secure REST endpoints with automatic scaling, without managing infrastructure.
Evaluation and metrics
Measure your model's response quality, coherence, groundedness and safety before production.
Integrated Azure AI Search
Connect knowledge bases to implement RAG (Retrieval-Augmented Generation) directly from the portal.
Content safety
Configurable content filters (hate, violence, sexual, self-harm) applied to the model's inputs and outputs.
Microsoft Foundry organizes resources in a two-level hierarchy.
Hub
Top-level organizational container. Defines the Azure subscription, region and shared connections (Azure OpenAI, Azure AI Search, Storage). A hub can group multiple projects for the team.
Project
Working workspace for a specific use case. Contains the project assets: model deployments, search indexes, datasets, evaluations and workflows.
Exam key point
Hub = organization / security. Project = workspace for the development team. Connections configured in the Hub are available to all projects it contains.
Chat and completion interface to explore models without code. Lets you switch between models, adjust temperature/top_p, configure system prompts and compare responses side by side.
Library of models available for deployment. Filter by task, provider, benchmark or license. Includes model cards with parameters, capabilities and pricing.
Management of models deployed as REST endpoints. Shows the deployment name, model version, type (Standard / PTU / Batch) and status.
Tools to measure response quality. Evaluates coherence, relevance, groundedness (backed by sources) and content safety with automated metrics.
Visual editor for AI orchestration workflows. It only exists in the classic portal, is no longer recommended for new development and retires on April 20, 2027; its replacement is Microsoft Agent Framework.
Safety filter configuration by severity for each harm category (hate, violence, sexual, self-harm) applied to prompts and completions.
Microsoft Foundry's model catalog brings together hundreds of models from multiple providers. It lets you compare, evaluate and deploy models from a unified interface — without managing your own infrastructure.
Microsoft
Phi-4, Phi-4-reasoning, Phi-4-mini, Phi-4-multimodal
Highly efficient Small Language Models (SLMs). Performance comparable to larger models on reasoning tasks, with lower cost and latency.
OpenAI
GPT-5 (and mini/nano variants), GPT-4.1 and GPT-4o (retiring), gpt-image, gpt-realtime, text-embedding-3
OpenAI models available through Azure with enterprise compliance, SLAs and guaranteed data privacy.
Meta
Llama 3.1, Llama 3, Llama 2 (7B, 13B, 70B)
Meta's open-source models. Available under a permissive license for commercial use. Managed deployment on Azure without managing servers.
Mistral AI
Mistral Large, Mistral Small, Mixtral 8x7B, Codestral
High-performance European models. Mistral Large competes with GPT models on reasoning and code benchmarks.
Cohere
Command R+, Command R, Embed (multilingual)
Specialized in RAG and enterprise search. Cohere's Embed is one of the best multilingual embedding models for vector search.
Catalog filters — how to choose a model
By task
text generation, embeddings, image generation, speech-to-text, translation
By provider
Microsoft, OpenAI, Meta, Mistral, Cohere, Hugging Face
By license
MIT, Apache 2.0, commercial, non-commercial use
By benchmark
MMLU, HumanEval, MT-Bench — objective performance comparison
Compare gpt-4o against Phi-3-mini and decide which to use based on the use case.
+1 more...
The Phi family are Microsoft's Small Language Models (SLMs) — small models optimized for high-quality reasoning with few parameters. Their differentiator: low cost, minimal latency, can run on edge devices.
| Model | Parameters | Best for |
|---|---|---|
| Phi-4 | 14B | Advanced mathematical reasoning, instruction following, complex tasks |
| Phi-4-reasoning | 14B | Step-by-step reasoning (math, logic) with a small model |
| Phi-4-multimodal | 5.6B | Text, image and audio in a single small model |
| Phi-4-mini | 3.8B | Edge devices, mobile, local inference without a dedicated GPU |
SLM vs LLM
LLM (Large Language Model) = GPT-5, Llama 70B — maximum capability, high cost. SLM (Small Language Model) = Phi-4-mini — efficient, economical, deployable at the edge. The Foundry catalog offers both to choose based on the use case.
1. Explore the catalog
In the Microsoft Foundry portal → Model catalog. Filter by task, provider or benchmark. Select the appropriate model.
2. Review the model card
View description, use cases, context limits, supported languages, license and available deployment options.
3. Choose the deployment type
Standard (pay-per-token), PTU (Provisioned Throughput Units — reserved capacity), Serverless API (billed by usage, no dedicated infrastructure).
4. Create the deployment
Assign a name to the deployment, select the model version, configure rate limits (tokens per minute). A REST endpoint with an API key is generated.
5. Test in the Playground
Verify the deployment in the integrated playground before connecting it to the application. Adjust the system prompt and generation parameters.
6. Integrate into the application
Use the generated endpoint with the Microsoft Foundry SDK, the Azure OpenAI SDK, or direct REST calls with the deployment API key.
Microsoft Foundry includes Azure OpenAI Service — they are not separate, competing services. Foundry is the complete platform; Azure OpenAI is the specific service for OpenAI's models.
| Aspect | Microsoft Foundry | Azure OpenAI Service |
|---|---|---|
| Scope | Complete unified GenAI platform | Specific service for OpenAI models (GPT, gpt-image, audio) |
| Models | All: Microsoft, OpenAI, Meta, Mistral, Cohere… | OpenAI models only |
| Portal | ai.azure.com — unified interface | Accessible from within AI Foundry or directly |
| Tools | Playground, agents, evaluation, fine-tuning, RAG | Playground, fine-tuning, OpenAI model deployments |
| Target audience | Teams building full GenAI apps | Direct integrations to GPT, gpt-image and audio models |
| Organization | Hubs + projects | Azure OpenAI resources |
Exam trap
"Microsoft Foundry contains Azure OpenAI Service" — they are not alternatives. If the scenario needs models from multiple providers, evaluation tools, or integrated RAG → Microsoft Foundry. If it only needs OpenAI models (GPT or gpt-image) → Azure OpenAI Service directly (or from within Foundry).
Did you get this topic?
Test what you just learned
A company wants to build a technical support assistant using generative AI. The team needs to: compare Microsoft and Meta models to choose the best one, implement RAG with their knowledge base, and evaluate response quality before production. Which Azure platform covers all these requirements?
Put what you learned to the test with original AI-901 exam questions.
Start AI-901 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.