Azure-AI_Foundry

Module 3: Model Catalog & Deployments — Overview

Introduction

Azure AI Foundry provides a Model Catalog — a centralized hub for discovering, evaluating, and deploying AI models from Microsoft and leading partners. Whether you need a powerful reasoning model, a lightweight edge model, or a specialized embedding model, the catalog gives you a single pane of glass to find, compare, and deploy.

Official Documentation: Model Catalog Overview

---

Available Model Families

Provider	Models	Strengths
OpenAI	GPT-4o, GPT-4o-mini, GPT-4.1, o3, o4-mini, DALL-E 3, Whisper, TTS, Embeddings	Best-in-class reasoning, multimodal, creative generation
Meta	Llama 3.1 (8B, 70B, 405B), Llama 3.2 (1B, 3B, 11B-Vision, 90B-Vision)	Open-weight, strong performance, permissive licensing
Mistral	Mistral Large, Mistral Small, Mistral NeMo, Codestral, Mixtral	Efficient inference, multilingual, strong code generation
Microsoft	Phi-4, Phi-3 (mini, small, medium), Phi-3.5 (MoE, Vision)	Small language models, edge-friendly, competitive benchmarks
Cohere	Command R, Command R+, Embed v3	Enterprise RAG, multilingual embeddings, grounded generation

---

Deployment Types

Azure AI Foundry supports multiple deployment options depending on the model and your workload:

1. Serverless API (Models as a Service — MaaS)

• No infrastructure to manage — pay per token
• Available for partner models (Meta Llama, Mistral, Cohere) and select Microsoft models
• Billed through Azure Marketplace
• Ideal for experimentation and variable workloads

2. Managed Compute (Models as a Platform — MaaP)

• Deploy open-weight models to dedicated VMs
• Full control over GPU SKU, scaling, and networking
• Best for fine-tuned models or when you need data isolation
• You manage the compute lifecycle and pay for VM hours

3. Azure OpenAI Service Deployment Tiers

Tier	Description	Use Case
Global Standard	Microsoft-managed, globally distributed	Default for most workloads, best availability
Standard	Regional deployment, shared capacity	Regional data residency requirements
Provisioned (PTU)	Reserved throughput units	Predictable latency, high-volume production
Global Provisioned	Provisioned with global routing	High-volume with global availability
Data Zone	Data stays within geographic zone	Compliance requirements (e.g., EU data zone)

---

Choosing a Deployment Type

Is it an Azure OpenAI model (GPT, DALL-E, Whisper)?
├── Yes → Azure OpenAI Deployment
│   ├── Variable traffic? → Global Standard (PAYG)
│   ├── Predictable high volume? → Provisioned (PTU)
│   └── Data residency needs? → Standard or Data Zone
│
└── No → Partner / Open Model
    ├── Quick experimentation? → Serverless API (MaaS)
    └── Need fine-tuning / data isolation? → Managed Compute

---

Quick Start: List Available Models with Python

from azure.ai.ml import MLClient
from azure.identity import DefaultAzureCredential

# Connect to your AI Foundry project
credential = DefaultAzureCredential()
ml_client = MLClient(
    credential=credential,
    subscription_id="your-subscription-id",
    resource_group_name="your-rg",
    workspace_name="your-ai-foundry-project"
)

# List models in the registry
models = ml_client.models.list(registry_name="azureml")
for model in models:
    print(f"{model.name}: {model.description[:80] if model.description else 'N/A'}")

---

Module Learning Objectives

By the end of this module, you will be able to:

1. Navigate the Model Catalog and evaluate models using benchmarks and model cards
2. Compare models across dimensions (quality, cost, latency, capabilities)
3. Deploy models using the appropriate deployment type for your scenario
4. Manage deployed models — monitor performance, handle quotas, plan for retirements

---

Module Lessons

Lesson	Topic
Lesson 1	Exploring and comparing models in the catalog
Lesson 2	Deploying models (OpenAI, Serverless, Managed Compute)
Lesson 3	Model management, monitoring, and cost strategies
Knowledge Check	15 scenario-based review questions

---

Tips

💡 Tip: Not every task needs GPT-4o. Start with a smaller model (GPT-4o-mini or Phi-4) and upgrade only if quality requires it — you’ll save significantly on cost.

💡 Tip: Serverless API deployments for partner models are billed through Azure Marketplace. Check pricing per 1M tokens before committing to a model.

💡 Tip: Use the “Benchmarks” tab in the catalog to compare models on standardized tasks before deploying — it saves you from running your own evaluations from scratch.