Microsoft Brings Qwen3.8-27B Coding Agents to Foundry Local

Microsoft has detailed a local coding setup that runs Qwen3.8-27B through Foundry Local and connects the model to GitHub Copilot. The approach keeps code and context on the developer’s machine, supports offline work, and offers more predictable inference costs than a cloud-only workflow.

With 4-bit quantization, Microsoft says the model is approximately 18 GB and can run on consumer Nvidia GPUs with at least 24 GB of VRAM. Foundry Local uses ONNX Runtime and combines INT4 weight handling, KV-cache quantization, embedding memory sharing, CPU offloading, prefix caching, and continuous batching. DFlash2 speculative decoding and GPU-specific kernel optimizations are intended to improve generation speed.

Microsoft compared Foundry Local with llama.cpp using the same 4-bit model on an RTX 5090 with 32 GB of VRAM and a DGX Spark with 128 GB of unified memory. In its test using real GitHub issues, Foundry Local generated faster at every measured prompt length, with a wider advantage beyond 131,000 tokens.

The setup requires Foundry Local CLI 0.13.0 or later, plus separately installed CUDA and compatible cuDNN runtimes. Microsoft also cautions that smaller local models may have less reasoning ability than advanced cloud models, making a hybrid workflow more suitable for complex tasks. Support for Macs and AMD GPUs is planned.