The AI Model Starter Kit

The fastest way to get started with building apps using any AI model.

Call any Foundry models — GPT, DeepSeek, Grok, Llama, and more — using stable OpenAI SDKs and the OpenAI Responses API. Production grade, enterprise friendly and secure with EntraID keyless authentication — no API keys, and zero secrets to manage.

Includes working examples in Python, TypeScript, C#, Java, and Go.

block-beta
    columns 5
    sdkTitle["OpenAI SDKs (Stable)"]:5
    py["🐍 Python"] ts["📘 TypeScript"] cs["🟪 C#"] java["☕ Java"] go["🩵 Go"]
    space:5
    space:1 arrow["⚡ OpenAI Responses API"]:3 space:1
    space:1 auth["🔐 EntraID Keyless Auth · Zero Secrets"]:3 space:1
    space:5
    foundryTitle["Microsoft Foundry Models"]:5
    deepseek["DeepSeek R1"] grok["xAI Grok"] llama["Meta Llama"] msft["Microsoft Phi"] oai["OpenAI GPT"]

    py --> arrow
    ts --> arrow
    cs --> arrow
    java --> arrow
    go --> arrow
    arrow --> foundryTitle

    style sdkTitle fill:#1a1a2e,color:#e0e0ff,font-weight:bold,stroke:#6c63ff,stroke-width:2px
    style foundryTitle fill:#1a1a2e,color:#e0e0ff,font-weight:bold,stroke:#e94560,stroke-width:2px
    style arrow fill:#16213e,color:#ffd460,font-weight:bold,stroke:#ffd460,stroke-width:2px
    style auth fill:#0d3320,color:#4ade80,font-weight:bold,stroke:#22c55e,stroke-width:2px
    style py fill:#3572A5,color:#fff,stroke:#4a8cc7,stroke-width:1px
    style ts fill:#3178c6,color:#fff,stroke:#4a9ae0,stroke-width:1px
    style cs fill:#9b4dca,color:#fff,stroke:#b366e0,stroke-width:1px
    style java fill:#e76f00,color:#fff,stroke:#f09030,stroke-width:1px
    style go fill:#00acd7,color:#fff,stroke:#33c4e8,stroke-width:1px
    style deepseek fill:#2563eb,color:#fff,stroke:#5b8def,stroke-width:1px
    style grok fill:#1d9bf0,color:#fff,stroke:#4db8f5,stroke-width:1px
    style llama fill:#6366f1,color:#fff,stroke:#8b8df5,stroke-width:1px
    style msft fill:#e94560,color:#fff,stroke:#f06e84,stroke-width:1px
    style oai fill:#10a37f,color:#fff,stroke:#3dbf9a,stroke-width:1px

Prerequisites

Azure Subscription
Azure Developer CLI
Azure CLI
One or more of: Python 3.9+, Node.js 18+, .NET 8+, Java 21+, Go 1.23+

Quick Start

# 1. Login to Azure
az login
azd auth login

# 2. Deploy DeepSeek-R1-0528 and gpt-4.1-mini to Microsoft Foundry
azd up

That's it! You now have a Microsoft Foundry account with both models deployed and ready to use.

Next Steps

Set up your environment

# 1. Set the project endpoint environment variable
$env:AZURE_AI_PROJECT_ENDPOINT = azd env get-value 'AZURE_AI_PROJECT_ENDPOINT'

# 2. Assign yourself the Azure AI Developer role
$userId = az ad signed-in-user show --query id -o tsv
$resourceId = "/subscriptions/$(az account show --query id -o tsv)/resourceGroups/rg-$(azd env get-value 'AZURE_ENV_NAME')/providers/Microsoft.CognitiveServices/accounts/$(azd env get-value 'AZURE_AI_FOUNDRY_NAME')"
az role assignment create --role "Azure AI Developer" --assignee $userId --scope $resourceId

Note: Role assignments can take up to 5 minutes to propagate. If you see 401/403 errors, wait a few minutes and retry.

Run an example

Each example calls both an OpenAI model (gpt-4.1-mini) and a non-OpenAI model (DeepSeek-R1-0528) using the Responses API.

Python

cd src/python
pip install -r requirements.txt
python responses_example.py

TypeScript

cd src/typescript
npm install
npx tsx responses_example.ts

C#

cd src/csharp
dotnet run

Java (requires Maven)

cd src/java
mvn -q compile exec:java

Go

cd src/go
go run .

The pattern (every language)

All five examples follow the same pattern:

Get an EntraID token via DefaultAzureCredential (scoped to https://ai.azure.com/.default)
Create a standard OpenAI client with base_url = project endpoint + /openai
Pass api-version as a query parameter (2025-11-15-preview)
Call the Responses API — works with any deployed model

Python code sample

import os
from azure.identity import DefaultAzureCredential
from openai import OpenAI

credential = DefaultAzureCredential()
token = credential.get_token("https://ai.azure.com/.default").token

endpoint = os.environ["AZURE_AI_PROJECT_ENDPOINT"]
client = OpenAI(
    base_url=endpoint.rstrip("/") + "/openai",
    api_key=token,
    default_query={"api-version": "2025-11-15-preview"},
)

# Works with any Foundry model — OpenAI, DeepSeek, etc.
response = client.responses.create(
    model="gpt-4.1-mini",  # or "DeepSeek-R1-0528"
    input="Explain quantum computing in simple terms",
    max_output_tokens=1000,
)
print(response.output_text)

The pattern (every language)

All five examples follow the same pattern:

Get an EntraID token via DefaultAzureCredential (scoped to https://ai.azure.com/.default)
Create a standard OpenAI client with base_url = project endpoint + /openai
Pass api-version as a query parameter (2025-11-15-preview)
Call the Responses API — works with any deployed model

Why the standard OpenAI SDK?

The Foundry project endpoint is OpenAI-compatible. By appending /openai and passing the api-version as a query parameter, you can use the standard openai library in any language — no azure-ai-projects, AzureOpenAI, or Azure.AI.OpenAI wrapper needed. This works for all model providers (OpenAI, DeepSeek, Meta, xAI, Microsoft, and more).

What This Template Includes

Core Infrastructure: Microsoft Foundry account + project with two model deployments
Optimal Configuration: GlobalStandard SKU, flexible region and model selection
Secure Authentication: EntraID (keyless) via DefaultAzureCredential — no API keys or secrets to manage
Multi-language examples: Python, TypeScript, C#, Java, and Go — all using their standard OpenAI SDK
Complete Documentation: Setup guide with troubleshooting tips

What You Get

DeepSeek-R1-0528 (default) — High-quality reasoning model, no registration required
Flexible model deployment — Deploy any Foundry model (OpenAI, DeepSeek, Meta, xAI, Microsoft, and more)
Microsoft Foundry project — Full project workspace for your development team
Standard OpenAI SDKs — Use each language's openai library directly, no Azure-specific wrappers
Responses API support — Modern API, cleaner than Chat Completions
Keyless authentication — Secure EntraID auth, no API keys to manage
Unique resource naming — No conflicts with existing resources

Template Structure

├── azure.yaml                 # azd configuration (includes postprovision hooks)
├── infra/
│   ├── main.bicep             # Main deployment template (subscription-scoped)
│   ├── main.parameters.json   # Deployment parameters
│   └── foundry.bicep          # Microsoft Foundry account and project
├── scripts/
│   ├── deploy-models.ps1      # Postprovision: deploy models via CLI (Windows)
│   ├── deploy-models.sh       # Postprovision: deploy models via CLI (Linux/macOS)
│   ├── list_models_with_responses_api_support.py  # Utility: list models supporting the Responses API
│   └── requirements.txt       # Dependencies for scripts (control plane)
├── src/
│   ├── python/
│   │   ├── responses_example.py         # Python example
│   │   └── requirements.txt
│   ├── typescript/
│   │   ├── responses_example.ts         # TypeScript example
│   │   ├── package.json
│   │   └── tsconfig.json
│   ├── csharp/
│   │   ├── Program.cs                   # C# example
│   │   └── FoundryModelsExample.csproj
│   ├── java/
│   │   ├── src/main/java/ResponsesExample.java  # Java example
│   │   └── pom.xml
│   └── go/
│       ├── main.go                      # Go example
│       └── go.mod
└── README.md

Language-specific SDK dependencies

Language	OpenAI SDK	Azure Identity
Python	`openai`	`azure-identity`
TypeScript	`openai`	`@azure/identity`
C#	`OpenAI` (NuGet)	`Azure.Identity`
Java	`com.openai:openai-java`	`com.azure:azure-identity`
Go	`github.com/openai/openai-go`	`github.com/Azure/azure-sdk-for-go/sdk/azidentity`

Configuration

Model deployments are created by postprovision hook scripts (scripts/deploy-models.ps1 / scripts/deploy-models.sh), which run automatically after azd up provisions the Foundry account. Configure models via azd env set before running azd up:

Variable	Default	Description
`AZURE_MODEL_NAME`	`DeepSeek-R1-0528`	Primary model to deploy
`AZURE_MODEL_FORMAT`	`DeepSeek`	Primary model format (`DeepSeek`, `Microsoft`, `OpenAI`, etc.)
`AZURE_MODEL_VERSION`	`1`	Primary model version
`AZURE_MODEL_DEPLOYMENT_NAME`	Same as model name	Custom deployment name for primary model
`AZURE_MODEL_SKU_NAME`	`GlobalStandard`	Primary model SKU tier
`AZURE_MODEL_SKU_CAPACITY`	`10`	Primary model capacity (tokens-per-minute in thousands)
`AZURE_MODEL_2_NAME`	`gpt-4.1-mini`	Second model to deploy (empty string to skip)
`AZURE_MODEL_2_FORMAT`	`OpenAI`	Second model format
`AZURE_MODEL_2_VERSION`	`2025-04-14`	Second model version
`AZURE_MODEL_2_DEPLOYMENT_NAME`	Same as model name	Custom deployment name for second model
`AZURE_MODEL_2_SKU_NAME`	`GlobalStandard`	Second model SKU tier
`AZURE_MODEL_2_SKU_CAPACITY`	`10`	Second model capacity

Example: DeepSeek-R1-0528 (Default)

azd up

Example: gpt-4.1-mini (OpenAI)

azd env set AZURE_MODEL_NAME "gpt-4.1-mini"
azd env set AZURE_MODEL_FORMAT "OpenAI"
azd env set AZURE_MODEL_VERSION "2025-04-14"
azd up

Example: Phi-4

azd env set AZURE_MODEL_NAME "Phi-4"
azd env set AZURE_MODEL_FORMAT "Microsoft"
azd up

Example: MAI-DS-R1

azd env set AZURE_MODEL_NAME "MAI-DS-R1"
azd env set AZURE_MODEL_FORMAT "Microsoft"
azd up

Outputs

After azd up, available via azd env get-values:

Output	Description
`AZURE_RESOURCE_GROUP`	Resource group name
`AZURE_AI_FOUNDRY_NAME`	Microsoft Foundry account name
`AZURE_AI_PROJECT_NAME`	Foundry project name
`AZURE_AI_PROJECT_ENDPOINT`	Project endpoint URL (use as `base_url` with `/openai` suffix)
`AZURE_MODEL_DEPLOYMENT_NAME`	Primary model deployment name
`AZURE_MODEL_2_DEPLOYMENT_NAME`	Second model deployment name

Common Commands

# Redeploy with changes
azd up

# Preview what would be deployed
azd provision --preview

# Debug deployment issues
azd up --debug

# Clean up everything
azd down --purge

Alternative Regions

azd env set AZURE_LOCATION eastus2
azd up

How Do I Find Which Foundry Models Support Responses and Where?

The scripts/list_models_with_responses_api_support.py utility queries the ARM control plane to discover which models support the Responses API in your subscription, along with their available regions. See the full list of supported models at the bottom of this page.

cd scripts
pip install -r requirements.txt

# List all models with Responses API support (OpenAI + non-OpenAI)
python list_models_with_responses_api_support.py

# Show per-region breakdown for models not available everywhere
python list_models_with_responses_api_support.py --locations

# List only non-OpenAI models (DeepSeek, Meta, xAI, etc.)
python list_models_with_responses_api_support.py --non-openai

# Show active model deployments across all Foundry accounts
python list_models_with_responses_api_support.py --deployments

# Combine flags
python list_models_with_responses_api_support.py --non-openai --locations

Note: The ARM control plane tags OpenAI-format models with capabilities.responses. For non-OpenAI models, capabilities.agentsV2 is used as a proxy — models with this capability support the Responses API at runtime. The default run combines both into a single unified list.

Known Issues

Issue	Summary	Workaround
ARM validation rejects non-OpenAI models	Bicep/ARM template deployment fails for non-OpenAI model formats (DeepSeek, Meta, Microsoft) with `DeploymentModelNotSupported`.	This template uses postprovision hook scripts that deploy models via Azure CLI, bypassing ARM validation. No action needed.
Java SDK `putQueryParam` bug	`openai-java` v4.21.0 `putQueryParam("api-version", ...)` silently drops the query parameter, causing `400: API version not supported` errors with EntraID auth.	Wait for an SDK fix. The Java example is included for reference but may not work until this is resolved.
API key endpoint limitation	The account-level API key endpoint (`/openai/v1`) does not support the Responses API for non-OpenAI models.	This template uses EntraID authentication with the project endpoint, which supports all models.

Troubleshooting

401 / 403 "Permission denied" — You need the Azure AI Developer role on the Foundry account. Run the role assignment command from Set up your environment. Role assignments can take up to 5 minutes to propagate.

"DeploymentModelNotSupported" during azd up — This shouldn't happen with the default configuration (models are deployed via CLI hooks, not Bicep). If you see this, ensure you're using the latest version of this template.

"Quota exceeded" — Check your subscription's quota for the selected model in the target region. Try a different region or model.

429 Rate limited — Increase AZURE_MODEL_SKU_CAPACITY or try a different model with more available capacity.

"api-version is required" / "Missing required query string parameter 'api-version'" — You're using the /openai path without passing api-version. Add api-version=2025-11-15-preview as a query parameter. All examples show how to do this.

"UnsupportedApiVersion" — The SDK is sending its default api-version (which isn't supported). Set api-version=2025-11-15-preview as a query parameter explicitly.

Java 400: API version not supported — This is a known SDK bug. The Java OpenAI SDK's putQueryParam doesn't correctly append the api-version. Wait for an SDK fix.

Need debug info? Run azd up --debug for detailed logs.

Why This Template?

Minimal setup — 2 commands instead of 20+
Any Foundry model — Access OpenAI, DeepSeek, Meta, xAI, Microsoft, and more
Any language — Python, TypeScript, C#, Java, and Go examples included
Modern API — Uses the Responses API, not legacy Chat Completions
Standard SDKs — Uses each language's OpenAI SDK directly, no Azure-specific wrappers
Production-ready — GlobalStandard SKU, EntraID auth, proper naming
Secure by default — Keyless EntraID authentication with Azure Identity
Easy cleanup — Remove everything with azd down --purge

Foundry Models with Responses API Support

All models currently available on Microsoft Foundry that support the Responses API, queried via the ARM control plane. Last updated: February 2026.

OpenAI

Model	Version(s)	Availability
codex-mini	2025-05-16	eastus2, swedencentral
gpt-4	0613	All except southeastasia
gpt-4-32k	0613	20/23 regions
gpt-4.1	2025-04-14	All regions
gpt-4.1-mini	2025-04-14	All regions
gpt-4.1-nano	2025-04-14	All regions
gpt-4o	2024-05-13, 2024-08-06, 2024-11-20	All regions
gpt-4o-mini	2024-07-18	All regions
gpt-4o-mini-realtime-preview	2024-12-17	eastus2, swedencentral
gpt-5-chat	2025-08-07, 2025-10-03	eastus2, swedencentral
gpt-5-mini	2025-08-07	All regions
gpt-5-nano	2025-08-07	All except canadacentral
gpt-5.1-chat	2025-11-13	8/23 regions
gpt-5.1-codex-mini	2025-11-13	8/23 regions
gpt-5.2-chat	2025-12-11	eastus2, swedencentral
o1	2024-12-17	All regions
o3-deep-research	2025-06-26	norwayeast, westus
o3-mini	2025-01-31	All regions
o4-mini	2025-04-16	All regions

Anthropic

Model	Version(s)	Availability
claude-haiku-4-5	20251001	eastus2, swedencentral
claude-opus-4-1	20250805	eastus2, swedencentral
claude-opus-4-5	20251101	eastus2, swedencentral
claude-opus-4-6	1	eastus2, swedencentral
claude-sonnet-4-5	20250929	eastus2, swedencentral
claude-sonnet-4-6	1	eastus2, swedencentral

DeepSeek

Model	Version(s)	Availability
DeepSeek-R1-0528	1	All except southeastasia
DeepSeek-V3-0324	1	All except southeastasia
DeepSeek-V3.1	1	All except southeastasia
DeepSeek-V3.2	1	All except southeastasia
DeepSeek-V3.2-Speciale	1	All except southeastasia

Microsoft

Model	Version(s)	Availability
MAI-DS-R1	1	All except southeastasia

OpenAI-OSS

Model	Version(s)	Availability
gpt-oss-120b	1	All except southeastasia

xAI

Model	Version(s)	Availability
grok-3	1	All except southeastasia
grok-3-mini	1	All except southeastasia
grok-4-fast-non-reasoning	1	All except southeastasia
grok-4-fast-reasoning	1	All except southeastasia

Region availability notes:

20/23 regions — gpt-4-32k is not available in japaneast, southeastasia, westeurope
8/23 regions — gpt-5.1-chat and gpt-5.1-codex-mini are available in australiaeast, canadaeast, eastus2, japaneast, koreacentral, swedencentral, switzerlandnorth, uksouth

Run python scripts/list_models_with_responses_api_support.py for the latest data from your own subscription.

Powered by Azure Developer CLI | Deploy any model on Microsoft Foundry

Name		Name	Last commit message	Last commit date
Latest commit History 19 Commits
.github		.github
infra		infra
scripts		scripts
src		src
.gitignore		.gitignore
README.md		README.md
ai-model-start.sln		ai-model-start.sln
azure.yaml		azure.yaml

Model	Version(s)	Availability
Llama-3.3-70B-Instruct	1, 2, 3, 4, 5	All except southeastasia
Llama-4-Maverick-17B-128E-Instruct-FP8	1	All except southeastasia

Azure-Samples/ai-model-start

Folders and files

Latest commit

History

Repository files navigation

The AI Model Starter Kit

Prerequisites

Quick Start

Next Steps

Set up your environment

Run an example

The pattern (every language)

Python code sample

The pattern (every language)

What This Template Includes

What You Get

Template Structure

Language-specific SDK dependencies

Configuration

Example: DeepSeek-R1-0528 (Default)

Example: gpt-4.1-mini (OpenAI)

Example: Phi-4

Example: MAI-DS-R1

Outputs

Common Commands

Alternative Regions

How Do I Find Which Foundry Models Support Responses and Where?

Known Issues

Troubleshooting

Why This Template?

Foundry Models with Responses API Support

OpenAI

Anthropic

DeepSeek

Meta

Microsoft

OpenAI-OSS

xAI

About

Resources

Code of conduct

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors 3

Uh oh!

Languages

Packages