Foundry Local brings Qwen coding to developer PCs

On November 5, Microsoft described how developers can run Qwen3.8-27B in Foundry Local and connect it to GitHub Copilot. The approach runs the coding model’s inference on the developer’s own hardware.

What hardware does the setup need?

Microsoft’s walkthrough calls for an NVIDIA GPU with at least 24 GB of VRAM and Foundry Local CLI 0.13.0 or later. CUDA and a compatible cuDNN runtime must be installed separately. Version 0.13.0 is a public preview, so features may change and rough edges are expected.

Microsoft says four-bit quantization reduces the model to about 18 GB. It describes support for up to 64K tokens of context with 24 GB of VRAM, and up to the model’s 262K-token maximum with 32 GB or more. Actual speed depends on the hardware and workload.

Connecting a local model to Copilot

After loading the model in Foundry Local, developers enter their local API endpoint in the GitHub Copilot app’s model-provider settings and select Qwen3.8-27B. Use the endpoint reported by your own machine rather than copying the example address in Microsoft’s screenshot.

Local inference means code and context need not be sent to a cloud model for that processing step. It does not guarantee that every connected tool is offline. Microsoft also notes that larger cloud models can remain useful for complex reasoning tasks.

What about Macs?

Foundry Local itself supports multiple platforms, but Microsoft presents this Qwen3.8-27B optimization around NVIDIA hardware. The company lists expanding this model support to Macs and AMD GPUs as future work. That is not confirmation that this particular setup already works on Macs.