microsoft-azure/pr_77/en/get-started/quickstart.md

Quickstart

Deploy a model from the Hugging Face collection on Microsoft Foundry, then call it from Python.

Before you begin

You need:

• An Azure subscription and a Microsoft Foundry project
• Cognitive Services Contributor access to create the deployment
• Managed compute quota for the accelerator that your model requires

Managed compute runs on dedicated GPU capacity and bills for as long as the deployment exists. Delete the deployments that you no longer need.

Deploy a model

Pick the Foundry portal for a first, guided deployment, or Python for a repeatable one. Both create the same managed compute deployment.

1. Open Microsoft Foundry and select your project.
2. Go to Discover → Models, and filter Collection by Hugging Face and Deployment options by Managed compute.
3. Select a model and choose Deploy.
4. Enter a deployment name, pick a deployment template and its accelerator, and select Deploy.

The deployment is ready when its state becomes Succeeded.

Install the dependencies and sign in:

pip install azure-identity "azure-mgmt-cognitiveservices>=15.0.0b2" openai
az login

Copy the Model ID, Deployment template ID, and Accelerator type from the model's deployment wizard in Foundry, then run:

from azure.identity import DefaultAzureCredential
from azure.mgmt.cognitiveservices import CognitiveServicesManagementClient

SUBSCRIPTION_ID = "<SUBSCRIPTION_ID>"
RESOURCE_GROUP = "<RESOURCE_GROUP>"
ACCOUNT_NAME = "<ACCOUNT_NAME>"
DEPLOYMENT_NAME = "qwen--qwen3-8-27b"

client = CognitiveServicesManagementClient(
    credential=DefaultAzureCredential(),
    subscription_id=SUBSCRIPTION_ID,
)

deployment = client.managed_compute_deployments.begin_create_or_update(
    resource_group_name=RESOURCE_GROUP,
    account_name=ACCOUNT_NAME,
    deployment_name=DEPLOYMENT_NAME,
    resource={
        "sku": {"name": "GlobalManagedCompute", "capacity": 1},
        "properties": {
            "model": "azureml://registries/azure-huggingface/models/qwen--qwen3.8-27b/versions/2",
            "deploymentTemplate": "azureml://registries/azure-huggingface/deploymenttemplates/qwen--qwen3-8-27b--256k-nvidia-h100/labels/latest",
            "acceleratorType": "H100_80GB",
            "versionUpgradeOption": "OnceNewDefaultVersionAvailable",
        },
    },
).result()

print(deployment.properties.provisioning_state)

The call returns when provisioning finishes, and prints Succeeded.

The deployment usually takes 15–20 minutes, but might take up to an hour whilst looking for instances when demand is high.

Call the deployment

Copy the endpoint and key from the deployment page, and use the deployment name as the model:

from openai import OpenAI

client = OpenAI(
    base_url="https://<ACCOUNT_NAME>.services.ai.azure.com/openai/v1",
    api_key="<API_KEY>",
)

response = client.chat.completions.create(
    model="<DEPLOYMENT_NAME>",
    messages=[{"role": "user", "content": "Write a one-sentence hello from Microsoft Foundry."}],
)

print(response.choices[0].message.content)

Some models are served through the Responses API instead, at /openai/v1/responses. Check the model page for the API that it supports.

Clean up

Delete the deployment to release its accelerators and stop billing, either from its page in Foundry or with:

client.managed_compute_deployments.begin_delete(
    resource_group_name=RESOURCE_GROUP,
    account_name=ACCOUNT_NAME,
    deployment_name=DEPLOYMENT_NAME,
).result()

Next steps

• Browse the catalog to pick a model.
• Turn reasoning on or off per request in Enable and disable reasoning.
• Connect the deployment to a harness such as Codex or Hermes Agent.