GitHub previews local-model routing in Copilot Fusion, not offline sessions
GitHub announced an opt-in experimental preview of intelligent local-model routing in Copilot Fusion on October 28. The feature lets Fusion choose between supported on-device and cloud models in Copilot CLI, the Copilot app and Visual Studio Code. It is distinct from the general availability of Fusion itself.
The managed download experience is limited to supported Windows devices and uses MAI Code 1.1 Flash through the Windows ML provider. GitHub describes separate settings for the CLI, app and VS Code. Device support and model readiness need to be checked before assuming local routing is available.
Once a supported model is ready, users select the Standard effort level in Fusion for local-or-cloud routing. They can also select a local model directly. The announcement does not establish that every prompt, completion or tool action will be processed on the device.
Copilot CLI can use manually configured supported local models on Windows, macOS and Linux through a local OpenAI-compatible server. Users manage the model download, runtime and device resources. Adding a model through bring-your-own-key configuration alone does not make it eligible for Fusion routing.
The Copilot app can directly select compatible local models through bring-your-own-key connections, but its intelligent routing currently supports managed Windows ML models only. Direct model selection and automated routing therefore have different compatibility limits.
Local inference does not make the session offline: GitHub says Fusion still makes routing decisions in the cloud. Local models may also have different safety protections from hosted models. This preview should not be treated as a guarantee of offline privacy, identical safeguards or zero overall session cost.