GitHub Copilot Fusion now supports selection of local models
By Steven Van ·
Opt-in preview routes Copilot requests between local and cloud models in the CLI, app and VS Code, with managed setup on Windows with an NVIDIA GPU.
Copilot Fusion can now route between on-device and cloud models, in an opt-in experimental preview. The routing works in GitHub Copilot CLI, the Copilot app and Visual Studio Code, and is meant to let Copilot use a device's compute without switching models by hand for each request. The details are in GitHub's documentation.
Managed setup on Windows
On supported Windows devices, MAI Code 1.1 Flash can be downloaded and configured through the Windows ML provider. Setup differs by surface:
- Copilot CLI: enable experimental mode and open /model to follow the setup prompts.
- Copilot app: configure the Windows ML provider under Settings > Model providers, enable Local model auto routing under Settings > Experiments, and start a new session. Windows ML local models is enabled by default on Windows.
- VS Code: enable chat.copilot.localModel.enabled.
Once the model is ready, selecting Standard in Copilot Fusion lets Copilot choose between local and cloud models. The local model can also be selected directly for a session. According to the docs, running MAI Code 1.1 Flash locally means Copilot can be used offline and avoids consuming GitHub AI Credits, with cloud models left for tasks that need more resources or higher performance.
Hardware requirements
The docs list these prerequisites for the in-app download: a Windows PC with an NVIDIA GPU (AMD is not supported), and llama.cpp v0.7.0 or later with LLAMACPP_SERVER_PATH set. Per model:
- MAI Code 1.1 Flash: 96 GB of available combined RAM, 60 GB of storage, llama.cpp build b11669 or later.
- Flint NVFP4: 48 GB of available combined RAM, 23 GB of storage, llama.cpp build b11803 or later.
Only MAI Code 1.1 Flash is available to Fusion Standard. A downloaded local model should not be used at the same time in sessions on different surfaces (CLI, app, VS Code), because each loads its own copy into memory and the machine is likely to run out of RAM.
Bring your own local model
The managed download is Windows-only, but Copilot CLI's local Fusion routing also works with manually configured models on Windows, macOS and Linux. MAI Code 1.1 Flash or Qwen3.8 Flash Next can be set up through a local OpenAI-compatible server. The user manages the model download and inference runtime, and the device needs enough resources to run it. The model must be configured as a supported local-model provider and experimental local routing enabled. Adding a model through BYOK does not automatically make it eligible for Fusion.
The Copilot app can also connect to and directly select compatible local models through BYOK, but its intelligent routing currently supports managed Windows ML models only.
Limits to know
Local inference does not make a session offline, since Copilot Fusion still makes routing decisions in the cloud. GitHub also notes that local models may have different safety protections from Copilot-provided cloud models.
