Microsoft brings Qwen3.8-27B agentic coding to developer PCs

Qwen3.8-27B targets local coding workflows

Microsoft Foundry identifies Qwen3.8-27B as its primary optimization target for local coding because it combines coding capabilities with a size suited to high-performance deployment on local hardware. With 4-bit quantization, the model can be reduced to approximately 18 GB, allowing it to run on consumer GPUs with 24 GB or more of VRAM, according to Microsoft.

The source lists results for four published benchmarks: 73.0 on Terminal Bench 2.1 for agentic terminal coding, 61.7 on SWE-bench Pro for agentic coding, 42.3 on NL2Repo-Bench for repository-level code generation, and 42.2 on DeepSWE 1.1 for agentic coding.

Microsoft says local inference can help developers avoid sending sensitive context over the network and continue working when an internet connection drops. It also presents local execution as a way to make inference costs more predictable, while acknowledging that smaller local models may have limited reasoning capabilities compared with advanced cloud models.

Foundry Local optimizes the full inference pipeline

Powered by ONNX Runtime, Foundry Local optimizes Qwen3.8-27B on NVIDIA GPUs across prefill, decoding, memory management, long-context processing and concurrent requests. The stack uses DFlash2 speculative decoding, which employs a lightweight drafter model to predict multiple tokens at once. It also includes CUDA graph fusions and kernel optimizations tuned for the latest NVIDIA SM121 GPU architecture.

To reduce memory pressure, the system combines INT4 weight handling, KV-cache quantization, embedding memory sharing and CPU offloading. Configurable, memory-aware inference is intended to balance memory use and performance across hardware configurations.

For repository-scale work and parallel requests, Foundry Local uses prefix caching and continuous batching. Microsoft says these features help keep the assistant responsive while developers explore a codebase, generate changes, run tools and iterate across tasks.

GitHub Copilot connection requires Foundry Local 0.13.0

Developers with an NVIDIA GPU containing at least 24 GB of VRAM can connect Foundry Local to GitHub Copilot through Bring Your Own Key, or BYOK. The setup requires Foundry Local CLI 0.13.0 or later. NVIDIA GPU execution also requires CUDA and compatible cuDNN runtimes to be installed separately, since Foundry Local does not download those dependencies.

The setup starts by installing the latest Foundry Local CLI with WinGet, then checking the version with foundry --version. Developers load the model with foundry model load qwen3.8-27b and retrieve the local API endpoint using foundry server status.

In the GitHub Copilot app, users open Settings, select Model providers, choose Add provider and select Foundry Local. They then enter the endpoint as the Base URL, add Qwen3.8-27B from Browse Catalog, and select the model after switching the provider to Foundry Local.

Microsoft plans broader hardware support

Microsoft says it plans to expand Qwen3.8-27B support to more devices, including Macs and AMD GPUs. It also plans to add more coding-optimized local models and deepen Foundry Local integration with GitHub Copilot.

The company describes local models as a fit for everyday coding and targeted subtasks, while cloud models remain a strong choice for complex, high-reasoning tasks. That positions Foundry Local as a hybrid option, allowing developers to keep routine coding on their own hardware and use cloud models when the task demands more reasoning capability.