AI 103 Part 3 LLM's

AI models

Microsoft Foundry — Model Catalog

Key models available today, and what each one is built for

11,000+ models available

What you can build

💬

Chat apps

GPT-4o, Claude Sonnet, Llama 3.3, Phi-4

🧠

Reasoning & agents

GPT-5.2, Claude Opus 4.6, o3, DeepSeek-V3

💻

Code generation

GPT-5.1-codex, Claude Opus, Codestral

🖼️

Image generation

DALL-E 3, GPT-image-1.5, FLUX.2, SD 3.5

👁️

Vision / image analysis

GPT-4o, Claude (all), GPT-5.2

🎙️

Speech & audio

gpt-4o-mini-transcribe, gpt-4o-mini-tts

📄

Long documents

Claude Opus 4.6 (1M ctx), GPT-5.2

⚡

High-volume / cheap

Claude Haiku 4.5, Phi-4, GPT-4o-mini

Model Providers

🟢

OpenAI — GPT Family

Microsoft's flagship — deepest Azure integration, most widely used

Azure-native

Model	Best for	Capabilities
gpt-5.2	Complex reasoning, enterprise agents, multi-step tasks	ChatAgentsVisionCode
gpt-5.1-codex	Engineering-scale code generation, debugging, refactoring	CodeAgents
gpt-4o	Balanced chat + vision + speed — great everyday workhorse	ChatVisionFast
gpt-4o-mini	Low-cost, high-volume tasks — same API as gpt-4o	ChatFastVision
o3	Deep reasoning — math, science, complex problem solving	ReasoningCode
gpt-image-1.5	Text-to-image generation, inpainting, image editing	Image gen
dall-e-3	High-quality image generation from text prompts	Image gen
gpt-4o-mini-transcribe	Fast, accurate speech-to-text transcription	SpeechFast
gpt-4o-mini-tts	Natural text-to-speech with multiple voice options	TTS

🟠

Anthropic — Claude Family

Best-in-class reasoning, coding, and 1M token context

Available in Foundry

Model	Best for	Capabilities
claude-opus-4-6	Most intelligent — complex agents, financial analysis, 1M token documents	AgentsLong docsCodeVision
claude-sonnet-4-6	Best balance of intelligence and speed — ideal for production apps	ChatAgentsVisionFast
claude-haiku-4-5	Fastest and cheapest Claude — high-volume tasks, real-time responses	ChatFast

🔵

Microsoft — Phi Family

Small, efficient models — great for edge, on-device, and cost-sensitive workloads

Microsoft-built

Model	Best for	Capabilities
phi-4	Punches above its size — strong reasoning at small model cost	ChatCodeFast
phi-4-mini	On-device and edge AI — runs locally with minimal resources	ChatEdge

🔷

Meta — Llama Family

Open-weight models — fully customizable, fine-tunable, self-hostable on Azure compute

Open-weight

Model	Best for	Capabilities
llama-3.3-70b	Best open-weight generalist — strong multilingual, fine-tunable	ChatCodeAgents
llama-3.2-vision	Open-weight model with image understanding capability	ChatVision

🟣

Specialized Models

Purpose-built for image generation, code, and research

Multi-vendor

Model	Best for	Capabilities
flux-2-pro	High-quality image generation — photorealistic, artistic styles	Image gen
stable-diffusion-3.5	Image + text input, fine-tunable image generation	Image genVision input
mistral-codestral	Code-specialized — fast autocomplete, fill-in-the-middle	CodeFast
deepseek-v3	Strong reasoning and coding — competitive with GPT-4 class	ReasoningCodeChat

💡 The Foundry advantage — all models share the same Azure AI inference API.
Switch from GPT-4o to Claude Sonnet by changing one line: model: "claude-sonnet-4-6"
No rewrites. No new SDK. Benchmark any model against your own data before committing.

Hugging face

Hugging Face is actually the biggest single source of models in Foundry. Here's the full picture:

The scale: Over 11,000 of the most popular and downloaded open-source models from the Hugging Face Hub are available in Foundry's model catalog — all with verified, security-scanned weights deployable to managed endpoints with one click. Microsoft Learn The partnership: Microsoft and Hugging Face have an expanded collaboration that puts new models in Foundry on the same day they land on the Hugging Face Hub — day-zero access — plus customized fine-tuned variants of trending models. Microsoft Learn Gated models too: Foundry now also supports Hugging Face's gated models — models behind an access boundary requiring approval from the model owner. You connect your Hugging Face access token to Foundry once, and it handles secure download and deployment automatically. CBT Nuggets Run locally: Foundry Local lets you convert any Hugging Face model (Safetensors or PyTorch) to ONNX using Olive and run it on your own device — fully offline, no Azure subscription needed.

The models can be invoked with API endpoints but also invoked from Foundry local

1 Install the tool

winget install Microsoft.FoundryLocal

List models

foundry model list

Download and run model

foundry model run phi-4-mini

and in second terminal get the endpoint

foundry service status

This gives you the local port, something like http://localhost:5273. Never hardcode this port — always fetch it at runtime since it's dynamic.

Use it in code

 using Microsoft.AI.Foundry.Local;

var manager = await FoundryLocalManager.StartModelAsync("phi-4-mini");
var client = new ChatClient(
    model: manager.GetModelId("phi-4-mini"),
    new ApiKeyCredential(manager.ApiKey),
    new OpenAIClientOptions { Endpoint = new Uri(manager.Endpoint) }
);

var response = await client.CompleteChatAsync("What is MCP?");
Console.WriteLine(response.Value.Content[0].Text);

Links Microsoft Foundry Local: Run AI Models On Your Device Building AI Apps with the Foundry Local C# SDK Example WPF Contoso medical service

Publishing an agent in Foundry