Loading page…
The opt-in preview works through Fusion Standard. Windows gets managed setup, while command-line users can configure supported models across three operating systems.
Listen to this story
GitHub’s October 28, 2026 experimental preview adds automatic routing between supported local and cloud models in Copilot Fusion, so developers can use device compute without choosing a different model for every request. The practical reach is limited: local models must be configured and eligible, and Fusion still makes routing decisions in the cloud, so on-device inference does not work offline. The update gives developers another way to manage model choice and compute, but local setup requirements and model-specific safety differences remain important constraints.
Managed Windows devices can configure MAI Code 1.1 Flash through Windows ML; CLI users can also connect MAI Code 1.1 Flash or Qwen3.8 Flash Next through a local OpenAI-compatible server.
Copilot app users can select compatible local models directly, but automatic local routing in the app currently supports only managed Windows ML models.
Fusion High and Max may use extra review models and AI credits; paid-plan users get a 30% discount on Fusion model costs through November 30, 2026.
GitHub released an opt-in experimental preview on October 28, 2026, letting Copilot Fusion choose between supported on-device and cloud models. Developers can use their computer’s processing power without switching models for each request. The update reaches Copilot CLI, the Copilot app and Visual Studio Code, but local execution does not make a session offline.
Local routing works through Fusion Standard. Once an eligible model is configured and experimental routing is enabled, Fusion can choose where to run a request. Adding or discovering a model through a bring-your-own-key connection does not automatically make it eligible.
Fusion itself is generally available in VS Code, Copilot CLI, the Copilot app and JetBrains IDEs. Selecting it delegates model choice rather than locking in one model. Standard uses one model per task; High and Max can bring in additional models.
In a cascade, an efficient model drafts a solution and a judge accepts it or escalates to a stronger model. Critique gathers parallel reviews before the original model revises its response. Both take longer than a single-model answer because they add review calls.
There is no separate Fusion fee: each model used is billed at its standard rate. High and Max can therefore consume more AI credits than Standard. Paid-plan users receive a 30% discount on Fusion model costs through November 30, 2026.
A discarded answer is not necessarily a clean rollback. If a draft has already edited files in the workspace, Fusion does not automatically undo those changes when it rejects the draft. GitHub tells developers to review changes before committing them.
Loading discussion...
Join the conversation
Explain whether offline access would matter to your choice.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.