Adding LLM models to the GitHub Copilot app

You can configure the GitHub Copilot app to include models from an LLM provider of your choice—using Bring Your Own Key (BYOK)—in addition to the GitHub-provided models. You can set up your model provider when you first open the app or later in app settings.

Windows users with an NVIDIA GPU and sufficient memory can install the MAI Code 1.1 Flash and Flint NVFP4 models locally on their device, directly from within the GitHub Copilot app.

You must sign in with a GitHub account to use the app, but you do not need a Copilot plan if you use your own model provider. If you do have a Copilot plan, you can use models from your own model provider and GitHub-provided models in the same app.

Adding models using BYOK

You can use the BYOK method to add models to the GitHub Copilot app.

Supported providers

The GitHub Copilot app supports these BYOK model providers:

• OpenAI
• Azure OpenAI
• Microsoft Foundry
• Anthropic
• Ollama
• Foundry Local
• LM Studio
• Any OpenAI-compatible HTTP endpoint

Prerequisite

You must have an API key to add a model provider that requires authentication.

Set up a model provider

If you are setting up the app for the first time and do not have a Copilot plan, you can choose to set up your own model provider during onboarding.

To add or update a model provider later:

1. Click to open the Settings dialog.
2. In the sidebar, click Model providers.
3. Click Add provider.
4. Select your provider.
5. Enter the provider details. This varies by provider but may include the display name, base URL, and API key.
6. Click Add provider to save the provider.

After you add a provider, its models appear in the model picker alongside GitHub-hosted models. You can select the model and use it in a session. For more information, see GitHub Copilot 앱에서 에이전트 세션 사용하기.

Provider credentials are stored in the system credential store and are never displayed in the UI.

Installing a local model from the model picker

If you are using Windows and your computer meets the following prerequisites, you can download and install the following models locally on your device from within the app's settings:

• MAI Code 1.1 Flash
• Flint NVFP4

Prerequisites

You must have:

• A Windows PC

• An NVIDIA GPU (AMD is not supported)

• Minimum available combined RAM per model:

  • MAI Code 1.1 Flash:  96 GB
  • Flint NVFP4:  48 GB

• Minimum available storage per model:

  • MAI Code 1.1 Flash:  60 GB
  • Flint NVFP4:  23 GB

• llama.cpp v0.7.0 or later installed and LLAMACPP_SERVER_PATH set to its path. Required build version:

  • MAI Code 1.1 Flash:  b11669 or later
  • Flint NVFP4:  b11803 or later

  See Installing llama.cpp for Windows.

Benefits of using MAI Code 1.1 Flash locally

MAI Code 1.1 Flash is a coding-optimized model that you can run locally on your Windows PC. By using a model that's running locally, you can use Copilot offline and avoid consuming GitHub AI Credits.

After installing MAI Code 1.1 Flash, you can select it from the model picker, just like any other model. The model also becomes available for Copilot Fusion's Fusion Standard effort level to use. When you set the model selection to Copilot Fusion with Fusion Standard, MAI Code 1.1 Flash will be automatically chosen to work on appropriate tasks, leaving cloud-based models for tasks that require more resources or higher performance.

Installing llama.cpp for Windows

To install llama.cpp on Windows with the CUDA build:

1. In your browser, go to the releases page of the llama.cpp repository: https://github.com/ggml-org/llama.cpp/releases.

2. Choose the release you want to install—for example, b11803.

3. Click the links to download both files for the Windows CUDA 13.4 ARM64 build:

   • Windows arm64 (CUDA 13)
   • CUDA 13.4 DLLs

4. Open Command Prompt and create the installation folder:

   Shell

   mkdir "%LOCALAPPDATA%\Programs\llama.cpp\BUILD-cuda-13.4-arm64"
   

   Replacing BUILD the the build number—for example, b11803.

5. Extract all files from both ZIP files directly into this folder. If you use Extract All in File Explorer, change the destination to this folder, because by default File Explorer creates a separate subfolder for each ZIP file. When you're done, the folder should contain a large number of files, including llama-server.exe, cudart64_13.dll, cublas64_13.dll, and cublasLt64_13.dll, with no subfolders.

6. Set the LLAMACPP_SERVER_PATH environment variable:

   Shell

   setx LLAMACPP_SERVER_PATH "%LOCALAPPDATA%\Programs\llama.cpp\BUILD-cuda-13.4-arm64\llama-server.exe"
   

   Replacing BUILD the the build number.

7. Close Command Prompt and open a new Command Prompt window. If the Copilot app or VS Code is already open, restart it. If Copilot CLI is running, exit it and start it again from a new terminal window.

8. To check the installation, run:

   Shell

   "%LLAMACPP_SERVER_PATH%" --list-devices
   

Installing a local model from within the Copilot app

1. Click to open the Settings dialog.

2. In the sidebar, click Model providers.

3. Click Add provider.

4. Select Windows ML.

   If your local machine meets the prerequisites, the supported local models will be listed under "Available models."

5. Click Download for the model you want to install.

6. Use the Maximum % system memory to use slider to cap the amount of system memory that can be used by Windows ML models.

7. Turn on Restart models on app restart to automatically restart downloaded Windows ML models when you restart the Copilot app.

8. Alter the model-specific settings as required.

Using a downloaded local model

After the download is complete, the local model will be available for use in your sessions.

You can either:

• Select it from the model picker, the same as any other available model. See Choosing a model in the Copilot app.
• MAI Code 1.1 Flash only: Choose Copilot Fusion, then Fusion Standard effort level in the model picker, and Copilot will automatically use this model when appropriate.
• Assign it to a custom agent for focused tasks. You can then prompt Copilot to use that custom agent for part of a task, while keeping your main session on a different model in the cloud. See 사용자 지정 에이전트 만들기 및 사용 GitHub Copilot CLI.