This guide explains how to run gpt-oss-120b on Azure AI Foundry, and how to run gpt-oss-20b using Foundry Local.

• Run gpt-oss-120b on Azure OpenAI
• Run gpt-oss-20b on Foundry Local
• Useful Links

This repo contains Python sample code for interacting with the gpt-oss-120b model deployed on Azure AI Foundry using Chat Completions. The gpt-oss models are OpenAI's open weight models that provide transparent access to its reasoning process.

• Python 3.7 or higher
• Azure subscription
• Azure CLI (optional, for deployment via command line)

Follow these steps to deploy and use the gpt-oss-120b model:

Deploy an Azure AI Foundry Project if you don't already have one available.

According to Microsoft Learn Docs, gpt-oss is available in all regions. I personally tested this using a Foundry Project in UK South.

📖 Detailed instructions: Create Azure AI Foundry Projects

Deploy the gpt-oss-120b model using one of the following methods:

1. Navigate to your Azure AI Foundry Project
2. Go to the Model catalog
3. Search for and select "gpt-oss-120b"
4. Click "Deploy" and follow the deployment wizard

az cognitiveservices account deployment create \
  --resource-group <your-resource-group> \
  --name <foundry-resource-name> \
  --deployment-name "gpt-oss-120b" \
  --model-name gpt-oss-120b \
  --model-version 1 \
  --model-format "OpenAI-OSS" \
  --sku-name GlobalStandard \
  --sku-capacity 1

Replace <your-resource-group> and <foundry-resource-name> with your actual values.

1. Clone this repository:

   git clone https://github.com/guygregory/gpt-oss.git
   cd gpt-oss

2. Install dependencies:

   pip install -r requirements.txt

3. Configure environment variables:

   cp .env.sample .env

4. Update the .env file with values from your Azure AI Foundry Project:

   • AZURE_OPENAI_API_ENDPOINT: Your foundry endpoint (format: https://<FOUNDRY_RESOURCE_NAME>.openai.azure.com/)
   • AZURE_OPENAI_V1_API_ENDPOINT: Your foundry v1 endpoint (format: https://<FOUNDRY_RESOURCE_NAME>.openai.azure.com/openai/v1/)
   • AZURE_OPENAI_API_KEY: Your API key (found under "Keys and Endpoint" in your project)
   • AZURE_OPENAI_API_MODEL: Your deployment name (default: gpt-oss-120b)

Choose one of the Python samples:

python chat-gradio-aoai.py

python chat-basic-aoai.py

python chat-basic-aoai-v1.py

Starting in May 2025, you can now opt in to our next generation of v1 Azure OpenAI APIs which add support for:

• Ongoing access to the latest features with no need to update api-version each month.
• OpenAI client support with minimal code changes to swap between OpenAI and Azure OpenAI when using key-based authentication.

Code samples have been provided for both the v1 API Preview, and also the older API versions. The v1 API Preview samples have a v1.py suffix to distinguish them.

If you want the latest features, I would recommend using the v1 API Preview, with the api-version set to preview. If you need a stable, GA version, and don't need the latest features, then you can use the older API. At time of writing, the latest GA API release is 2024-10-21.

Azure OpenAI in Azure AI Foundry Models API lifecycle

1. Authentication Error: Verify your API key and endpoint URL are correct
2. Model Not Found: Ensure the deployment name matches your Azure deployment
3. Region Not Supported: Verify you deployed in one of the supported regions
4. Rate Limiting: The gpt-oss model has usage quotas; check your deployment capacity

• Check the Azure AI Foundry documentation
• Review Azure OpenAI service logs in the Azure portal
• Verify your deployment status in the Azure AI Foundry project

If you want to experiment with the gpt-oss-20b model locally without using Azure, you can use Foundry Local.

• NVIDIA GPU with at least 16 GB VRAM
• Foundry Local version 0.6.87 or above

Check your version with:

winget list --id Microsoft.FoundryLocal

Update to the latest version (if needed):

winget upgrade Microsoft.FoundryLocal

foundry model run gpt-oss-20b

💡 Note: This will download several GBs of model weights if not already cached.

• Documentation for running gpt-oss-20b on Foundry Local can be found here
• Foundry Local is ideal for local development, offline prototyping, or sandbox testing
• If you get the error Exception: Model <gpt-oss-20b> was not found in the catalog or local cache., check your PC meets the 16GB NVIDIA requirement, and you are running version 0.6.87 or above. Reboot your PC after upgrading.
• Works well for basic chat and reasoning tasks if your hardware meets the requirements

Here are some useful links related to gpt-oss:

• OpenAI Announcement: OpenAI Open Models
• Microsoft Announcement: OpenAI's Open-Source Model gpt-oss on Azure AI Foundry and Windows AI Foundry

• Azure Region Availability: gpt-oss Model Availability
• Run gpt-oss-20b on Foundry Local: Get Started Guide
• Model Card: gpt-oss Model Card (PDF)

• Code Samples: This Repository
• HuggingFace: gpt-oss Collection
• Ollama: gpt-oss Library
• GitHub: OpenAI gpt-oss Repository
• Playground (OpenAI): gpt-oss.com

This project is licensed under the MIT License - see the LICENSE file for details.