What changed
Microsoft showed how to connect the existing Qwen3.8-27B model to the GitHub Copilot app. Foundry Local runs Qwen on your computer and gives Copilot a local address to send requests to. Qwen generates the model's response on that machine. [1]
Check your graphics memory first
Microsoft’s setup uses a smaller version of the model that takes about 18 GB and targets NVIDIA graphics cards with at least 24 GB of their own memory. It calls for Foundry Local’s command-line tool version 0.13.0 or newer, plus separately installed NVIDIA software called CUDA and cuDNN. The first load downloads the model if you do not already have it. [1] [2]
The guide uses a local address that Foundry Local reports when it starts. GitHub calls adding your own model provider Bring Your Own Key. Microsoft lists support for this optimized Qwen setup on Macs and AMD graphics as future work. [1] [3]
Our take
Use two copies of a small project and give your usual model and the local model the same request. For example, ask each to fix a sorting error, add one test, and leave unrelated files alone. Compare the changed files and run the test yourself. If one assistant also rewrites the page or changes a setting, that is useful to know before you give it a larger job.
Start with a project that contains no sensitive data. A model that generates responses locally is only one part of the setup. Check which tools the coding app can call and whether any of those tools connect elsewhere before using private code. We read the setup instructions and did not run the model or test the integration.