Foundry Local runs Qwen3.8-27B in GitHub Copilot on-device
Microsoft Foundry Local now serves Qwen3.8-27B to GitHub Copilot via BYOK on a 24GB NVIDIA GPU. On-device agentic coding, Windows only. Keep cloud for hard tasks.
Your coding agent can now run on the GPU under your desk, inside the Copilot app you already use. On November 5, Microsoft published a guide to running Qwen3.8-27B through Foundry Local and plugging it into GitHub Copilot as a bring-your-own-key model provider. On-device agentic coding with Foundry Local means no per-token bill and no code leaving the machine. It also means a model Microsoft itself says may reason less well than the cloud.
What actually happened
From Microsoft's Foundry blog post:
- Hardware. An NVIDIA GPU with at least 24 GB of VRAM. The 4-bit model is about 18 GB. 24 GB supports context up to 64K tokens; 32 GB or more goes toward the model's 262K maximum.
- Software. Foundry Local CLI 0.13.0 or later (
winget install Microsoft.FoundryLocal). CUDA and cuDNN must be installed separately; Foundry Local does not download them. - Setup.
foundry model load qwen3.8-27b, thenfoundry server statusfor the endpoint. In the Copilot app, add Foundry Local under Model providers, paste the endpoint as the Base URL and pick Qwen3.8-27B. - Benchmarks. Microsoft cites the Qwen model card: 73.0 on Terminal Bench 2.1 and 61.7 on SWE-bench Pro.
- Speed. Microsoft claims faster generation than llama.cpp on an RTX 5090 and a DGX Spark, but publishes the results only as charts.
- Limits. Windows only for now. Mac and AMD GPU support are listed as "next." Microsoft recommends a hybrid workflow, since local models "may have limited reasoning capabilities."
The post does not state which Copilot plans allow BYOK; it links to GitHub's docs. The 0.13.0 release link is labeled as a CLI preview.
Why it matters for your business
Two kinds of shops should care. One has code it cannot send out: client contracts, regulated data, or a buyer who asks where the source goes. The other has a developer burning through cloud tokens on routine edits.
For both, the option is now practical. Compare the price of a 24 GB card against your monthly API bill per developer, and note that the setup is four commands. But do not move everything. Route boilerplate, tests, refactors and anything touching sensitive code to the local model. Keep architecture questions and hard multi-file bugs on a frontier model in the cloud. Copilot's provider switch makes that a dropdown, not a migration.
Key takeaways
- Foundry Local serves Qwen3.8-27B to GitHub Copilot as a BYOK provider, per Microsoft's November 5 post
- Needs an NVIDIA GPU with 24 GB VRAM (64K context) or 32 GB+ for longer context
- Windows only; Mac and AMD support are not shipped yet
- Microsoft's speed claims are chart-only; benchmark it on your own repo
- Use local for routine and sensitive code; keep hard reasoning on cloud models
Want AI coding without your source leaving the building? Rush Commerce sets up hybrid local-and-cloud dev workflows that route each task to the right model. Run the ROI calculator on your token spend, or tell us what your team builds.
Sources: Microsoft Foundry Blog.
- #foundry-local
- #github-copilot
- #qwen
- #local-llm
- #agentic-coding
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Gemini Omni 1.2 Flash: $0.07/sec, 15-second product video
Gemini Omni 1.2 Flash is generally available at $0.07 per second of 720p, 30% cheaper, with 15-second clips and multi-turn edits. Here is the product-video math.
Read itOpenAI Images 2.6: GIFs and SVGs, but no Batch API
OpenAI Images 2.6 adds Flare and Sunburst models with GIF and SVG output, but no Batch API. Keep bulk catalog jobs on 2.5 Batch at half the output price.
Read it