GitHub Copilot Fusion Introduces Intelligent Local Model Routing in Preview

GitHub has launched an experimental preview of intelligent routing for local models within GitHub Copilot Fusion. This update allows the Copilot CLI, Copilot app, and Visual Studio Code to automatically determine whether to utilize an on-device local model or a cloud-based model for specific tasks, reducing the need for manual switching.

Key Features of the Experimental Preview

• Intelligent Routing: When the 'Standard' option is selected in Copilot Fusion, the system dynamically routes requests between eligible local and cloud models.
• Managed Model Support: On Windows devices, users can download and configure MAI Code 1.1 Flash through the Windows ML provider.
• Manual Configuration: Users on Windows, macOS, and Linux can connect manually configured models, such as MAI Code 1.1 Flash or Qwen3.8 Flash Next, via an OpenAI-compatible server.

Configuration Guidelines

• Visual Studio Code: Enable the `chat.copilot.localModel.enabled` setting to begin.
• Copilot CLI: Enable experimental mode and navigate to `/model` to follow the setup prompts.
• Copilot App: Configure the Windows ML provider in settings and enable 'Local model auto routing' under the experiments menu.

Important Considerations

While this feature leverages local hardware, it is not a fully offline experience. Copilot Fusion still processes routing decisions in the cloud. Additionally, users should be aware that local models may carry different safety protections compared to standard GitHub-provided cloud models. Users are responsible for managing the inference runtime and ensuring their hardware meets the necessary resource requirements. For detailed configuration steps, please refer to the official GitHub Copilot documentation on adding LLM models.