The Pulse
Foundry Local Connects Qwen3.8-27B to GitHub Copilot for Local Coding, Microsoft Says
Foundry Local can run Qwen3.8-27B on developer hardware and connect it to GitHub Copilot, keeping code and context on the user’s machine, Microsoft says. The company describes the setup, its GPU optimizations and a comparison with llama.cpp
AI.info Team · Written with AI · fact-checked ·
Foundry Local can run Qwen3.8-27B on developer hardware and connect it to GitHub Copilot, bringing agentic coding into a local workflow, Microsoft says. The Microsoft Foundry blog published setup instructions on Nov. 5, 2026. Local inference keeps code and context on the user’s machine, can continue working offline and offers predictable inference costs, according to Microsoft.
Connecting Foundry Local to GitHub Copilot
The setup uses GitHub Copilot’s Bring Your Own Key option, with Foundry Local selected as a model provider. It requires Foundry Local CLI 0.13.0 or later and an NVIDIA GPU with at least 24 GB of VRAM. For NVIDIA GPU execution, users must separately install CUDA and compatible cuDNN runtimes; Foundry Local does not download them. The runtime library paths also need to be configured, including adding DLL directories to PATH on Windows.
After installing or updating the CLI, users can check its version, load Qwen3.8-27B and retrieve the local API endpoint with the Foundry Local server status command. Loading the model triggers a download if it is not already on the device. Developers then enter their machine’s endpoint in GitHub Copilot’s model-provider settings, add Qwen3.8-27B from Browse Catalog and choose it in the model picker. Users should use the endpoint returned on their own machine rather than copying the port shown in an image.
Memory and inference optimizations
Microsoft’s 4-bit quantization reduces Qwen3.8-27B to approximately 18 GB, making it runnable on consumer GPUs with 24 GB of VRAM and allowing context windows of up to 64K tokens. GPUs with 32 GB or more of VRAM can support longer context windows, up to the model’s maximum supported length of 262K tokens.
Powered by ONNX Runtime, Foundry Local optimizes the inference pipeline on Nvidia GPUs, including prefill, decoding, memory management, long-context processing and concurrent requests. The stated techniques include DFlash2 speculative decoding, which uses a lightweight drafter model to predict multiple tokens, as well as CUDA graph fusions and kernel optimizations tuned for the latest NVIDIA SM121 GPU architecture.
To reduce memory pressure, the stack also combines INT4 weight handling, KV-cache quantization, embedding memory sharing and CPU offloading. Configurable, memory-aware inference can balance memory use and performance across hardware configurations, Microsoft says. For work involving large repository contexts or multiple requests, the serving infrastructure uses prefix caching to reuse previously computed context and continuous batching to process concurrent workloads.
Speed advantage over llama.cpp, Microsoft reports
Microsoft compared Foundry Local with llama.cpp using the 4-bit Qwen3.8-27B model on an NVIDIA GeForce RTX 5090 with 32 GB of VRAM and an NVIDIA DGX Spark with 128 GB of unified memory. The evaluation used a dataset of real GitHub issues. Foundry Local delivered higher generation speed at every tested prompt length, Microsoft reports, with the advantage widening beyond 131K tokens.
The blog also lists Qwen3.8-27B benchmark scores of 73.0 on Terminal Bench 2.1 for agentic terminal coding, 61.7 on SWE-bench Pro, 42.3 on NL2Repo-Bench for repository-level code generation and 42.2 on DeepSWE 1.1.
Local and cloud models for different coding tasks
Microsoft cautions that local models may have more limited reasoning capabilities than advanced cloud models because of their smaller size. It describes cloud models as a strong choice for complex, high-reasoning work and says local models can handle everyday coding and targeted subtasks, allowing developers to use both types of models in a workflow.
Microsoft plans to expand Qwen3.8-27B support to more devices, including Macs and AMD GPUs, add more coding-optimized local models and deepen Foundry Local integration with GitHub Copilot.