Microsoft has added a way to use Qwen3.8-27B as a local coding model in GitHub Copilot through Foundry Local. Developers with an Nvidia GPU with at least 24 GB of VRAM can connect the model to the Copilot app and run it on their own hardware.
The setup requires Foundry Local CLI 0.13.0 or later. Developers must also install the required CUDA and compatible cuDNN runtimes separately. Microsoft says Foundry Local does not download those dependencies.
Microsoft says its 4-bit version of Qwen3.8-27B is about 18 GB. On a GPU with 24 GB of VRAM, it supports context windows of up to 64K tokens. GPUs with at least 32 GB of VRAM can support up to the model’s maximum of 262K tokens. Foundry Local uses ONNX Runtime and techniques including speculative decoding, memory management and caching to handle local requests.
Microsoft reports faster generation in its tests
Microsoft compared Foundry Local with llama.cpp using the 4-bit model on an Nvidia GeForce RTX 5090 with 32 GB of VRAM and an Nvidia DGX Spark with 128 GB of unified memory. In tests using real GitHub issues, the company says Foundry Local generated text faster at every prompt length it tested, with a wider lead beyond 131K tokens.
To use the model in the GitHub Copilot app, developers load it in Foundry Local, get the local API endpoint and add Foundry Local as a model provider in Copilot’s settings. They can then choose Qwen3.8-27B in the model picker. Microsoft says smaller local models may have limited reasoning capability compared with advanced cloud models on complex tasks.
Microsoft says it plans to extend Qwen3.8-27B support to more devices, including Macs and AMD GPUs, add more coding models and deepen the Foundry Local integration with GitHub Copilot.
