Copilot Fusion previews local-model routing, not an offline mode
GitHub's Universe release adds experimental local and cloud model routing, with managed Windows setup and important cloud dependencies.
- #GitHub
- #Copilot
- #Universe
- #local-models
At Universe on October 28, GitHub announced experimental preview support for local-model selection in Copilot Fusion. Supported inference can run on a developer’s device while Fusion chooses between local and cloud models.
This is separate from Fusion’s general availability. It is also not a claim that Copilot now works entirely offline.
The supported paths are not identical
GitHub lists the Copilot CLI, Copilot app, and Visual Studio Code as surfaces for the experimental capability. The managed download and setup experience is Windows-specific, using MAI Code 1.1 Flash through the Windows ML provider on supported devices.
The CLI also supports manually configured compatible local models on Windows, macOS, and Linux through an OpenAI-compatible server. In that workflow, the developer is responsible for the model download, inference runtime, and available device resources.
Connecting a local provider is not the same as enabling automatic routing everywhere. The announcement distinguishes the app’s compatible-provider configuration from its intelligent local-routing experience, which currently uses the managed Windows flow. Teams should check the instructions for their particular client rather than assuming platform parity.
Local inference still has cloud dependencies
The most important limitation is explicit in GitHub’s changelog: Fusion’s routing decisions still happen in the cloud. A session can also use cloud models. Running some inference locally therefore does not establish that all task data stays on the device.
GitHub also cautions that local models may not have the same safety protections as cloud-provided models. Local execution is a deployment choice, not a replacement for permission controls or review.
For teams with offline or data-residency requirements, the relevant question is the complete data path: what goes to the router, which models can be selected, and what each client sends outside the device. This preview alone should not be treated as evidence that those requirements are met.
A useful pilot has measurable boundaries
Start with non-sensitive tasks on supported hardware. Compare end-to-end latency and output quality against the existing cloud workflow, including the time spent reviewing changes. Device resource use and cloud fallback matter as much as whether a local model successfully starts.
The opportunity is to use device compute where it helps while retaining cloud capacity for other work. The boundary is equally clear: the October 28 release, reviewed on October 31, 2026, remains an experimental hybrid workflow rather than an offline coding assistant.
Sources
- GitHub Copilot Fusion now supports selection of local models - GitHub Changelog, October 28, 2026
- Route and orchestrate models with GitHub Copilot Fusion - GitHub Changelog, October 28, 2026