GitHub Copilot Fusion now supports selection of local models

GitHub launched intelligent routing for local models into experimental preview for GitHub Copilot Fusion, allowing the system to alternate between on-device models and remote cloud endpoints. According to GitHub Changelog's report, the feature coordinates device compute across GitHub Copilot CLI, the GitHub Copilot app, and Visual Studio Code so developers avoid switching models manually for every request. On supported Windows systems, engineers can access managed installations through Windows ML, while cross-platform environments can point to local inference engines without leaving the Copilot workflow.

This breakdown examines the routing configuration, host prerequisites, and platform trade-offs for engineers integrating local machine learning with GitHub Copilot tools. The preview specifically targets developers seeking to preserve local hardware capacity on Windows, macOS, and Linux without giving up cloud fallback. It covers setup paths inside editor settings, command-line interfaces, model support parameters, and security boundaries that govern on-device execution.

What GitHub Copilot Fusion now supports selection shipped

GitHub Copilot Fusion shipped an automated dispatch layer that dynamically selects between local on-device models and remote cloud models based on task context. In the experimental preview, the runtime coordinates model invocation automatically when users designate the Standard profile inside Copilot Fusion. Instead of locking a project or terminal window to a single static endpoint, the system assesses incoming prompts to decide whether the workload can resolve on local silicon or requires higher-tier cloud inference. The functionality integrates directly across three client tools: GitHub Copilot CLI, the desktop Copilot app, and the Visual Studio Code extension environment.

The deployment introduces platform-native model delivery on Windows alongside generic server configuration for other operating systems. Windows hardware with Windows ML support provides an automated download and initialization pipeline for MAI Code 1.1 Flash. For cross-platform environments across macOS, Linux, and Windows, Copilot Fusion enables intelligent dispatch against local OpenAI-compatible inference servers running MAI Code 1.1 Flash or Qwen3.8 Flash Next. This expands upon the foundational architecture GitHub previously previewed on October 7th, shifting local machine learning from manual toggles into an orchestrated routing layer.

What changed for builders in GitHub Copilot Fusion now supports

Builders no longer need to execute repetitive manual model swaps across separate terminals or settings menus when moving between small routine edits and complex code refactors. Enabling the Standard routing profile in Copilot Fusion hands dynamic delegation over to the system while keeping the option open to pin an on-device model directly for dedicated local sessions. On Windows installations, Windows ML local models is activated by default, reducing friction for engineers adopting on-device acceleration without external wrapper scripts.

The update formalizes runtime requirements across client surfaces, distinguishing between fully managed runtime setups and bring-your-own-engine configurations.

For cross-platform developers using GitHub Copilot CLI, the change leverages recent local-model discovery additions while applying specific eligibility criteria. Simply exposing or discovering an endpoint through existing bring your own key (BYOK) settings does not qualify that model for automated Copilot Fusion dispatch. Teams must explicitly pair eligible model weights such as MAI Code 1.1 Flash or Qwen3.8 Flash Next through OpenAI-compatible server APIs and mark them as supported local providers to participate in automatic balancing.

How to install or upgrade GitHub Copilot Fusion now supports selection

Getting started with the experimental preview requires activating experimental options within each targeted developer interface. In Visual Studio Code, open user settings and turn on the chat.copilot.localModel.enabled configuration flag to expose on-device routing hooks. Inside the GitHub Copilot desktop app on Windows, navigate to Settings, open Model providers to configure the Windows ML provider, proceed to Experiments to enable Local model auto routing, and start a new workspace session to apply the changes.

In GitHub Copilot CLI, builders enable experimental mode directly in their environment and trigger setup prompts through the model interface. To switch or inspect active models within the interactive interface, invoke the model menu command:

/model

Once the underlying provider downloads or binds to the local host runtime, choose Standard inside Copilot Fusion to authorize automated switching between local execution and cloud engines. Developers who wish to bypass dynamic balancing can also select the installed local model directly from the model picker menu to force continuous on-device processing.

Gotchas and compatibility in GitHub Copilot Fusion now supports

Local execution does not transform GitHub Copilot into an offline development system. Copilot Fusion still routes requests through cloud services to make routing determinations before directing execution to the local runtime. Furthermore, local endpoints operate under different safety controls than standard GitHub Copilot cloud models. Development teams evaluating local execution must also verify that target hardware supplies sufficient memory and processor overhead to host MAI Code 1.1 Flash or Qwen3.8 Flash Next without stalling the primary development workstation.

Platform capabilities also diverge significantly between native operating system clients. The automated download and setup experience remains exclusively available on Windows machines using the Windows ML provider. While the desktop Copilot app allows connecting and selecting compatible local engines via BYOK, its intelligent auto-routing mechanism currently functions only with managed Windows ML installations. Developers on macOS and Linux must run an external OpenAI-compatible server, manage local engine downloads independently, and interact primarily through the GitHub Copilot CLI to leverage automated routing.

What to watch after GitHub Copilot Fusion now supports selection

Engineers adopting the preview should monitor system resource utilization during intense coding sessions as local inference processes scale. Because bring-your-own-model eligibility remains restricted to specific models like MAI Code 1.1 Flash and Qwen3.8 Flash Next, teams must watch how GitHub updates its eligible provider list for Copilot CLI routing. Developers managing their own inference servers should also check latency differences between local processing and cloud dispatch to ensure their local inference stack reliably yields speed advantages during routine operations.

Compatibility across client interfaces remains another critical area to track as the preview matures. The current gap between the desktop Copilot app, which restricts auto-routing to Windows ML, and Copilot CLI, which supports cross-platform OpenAI-compatible servers, points toward evolving client parity. Teams operating mixed OS environments must maintain separate onboarding workflows for Windows and Unix-like machines until uniform local configuration mechanisms deploy across all Copilot client applications.