GitHub Copilot has always been a cloud product. Every completion, every agentic task, every line of your proprietary code — it all hit a Microsoft server. That changes now. Microsoft’s Foundry Local, updated November 5, lets you load Qwen3.8-27B directly onto your GPU and wire it into Copilot through the Bring Your Own Key model picker. The prompts never leave your machine.
This is not a toy setup. Qwen3.8-27B scores 73.0 on Terminal Bench 2.1 and 61.7 on SWE-bench Pro — numbers that beat Claude Opus 4.6 Max on coding benchmarks, from a 27B open-weight model you can run locally for free.
How to Set It Up #
You need Foundry Local CLI version 0.13.0 or later. The process takes three steps:
foundry model load qwen3.8-27b
foundry server status
Once the server is running, open GitHub Copilot’s settings, navigate to Model providers → Add provider → Foundry Local, browse the catalog, and add Qwen3.8-27B. Switch to it in the model picker and Copilot starts sending requests to your local endpoint instead of Microsoft’s cloud.
One dependency note: CUDA and cuDNN must be installed separately. They are not bundled with Foundry Local, and the model will not load without them. Budget 10–15 minutes for dependency setup if you are starting clean.
The Model’s Actual Performance #
Qwen3.8-27B is not just capable — it is a meaningful generational jump over its predecessor. On DeepSWE 1.1, it went from 13.3 (Qwen3.6-27B) to 42.2, a more than 3x improvement in a single release. That benchmark measures autonomous debugging and repair, which is exactly what agentic coding agents need to do.
| Benchmark | Qwen3.8-27B |
|---|---|
| Terminal Bench 2.1 | 73.0 |
| SWE-bench Pro | 61.7 |
| DeepSWE 1.1 | 42.2 |
| LiveCodeBench v6 | 90.3 |
The model is licensed under Apache 2.0 on HuggingFace, so there are no commercial restrictions on using it in your workflow.
Why Local Wins for Agentic Tasks #
Foundry Local uses ONNX Runtime with DFlash2 speculative decoding, prefix caching, and continuous batching. Microsoft benchmarked it against llama.cpp on an RTX 5090 and found Foundry Local faster at every prompt length — with the gap widening significantly past 131K tokens.
That 131K crossover point matters. Agentic coding tasks — the kind where Copilot reads your entire repo, plans multi-file changes, and executes a sequence of steps — routinely hit those context lengths. Cloud Copilot is fast enough for autocomplete. For long-horizon agentic work, local inference with prefix caching compounds its advantage with every token you add. There is also the cost angle: cloud Copilot bills per token. Run the same workload locally and those costs disappear.
The Hardware Requirements Are Real #
The 24GB VRAM floor is non-negotiable. 4-bit quantization brings the active model size to roughly 18GB, but you need a 24GB card to load it reliably. That means RTX 3090, 4090, or 5090. Nothing below qualifies.
- NVIDIA only. AMD and Mac support are on the roadmap but not available today. Mac users can run other Foundry Local models via Metal, but Qwen3.8-27B specifically needs CUDA.
- CUDA and cuDNN are required separately. The model fails to load without them — do not skip this step.
- Context scales with VRAM. 64K context on 24GB, 262K context on 32GB or more.
When to Use Local vs Cloud Copilot #
If you have the GPU, there is now no compelling argument for sending proprietary code to the cloud for agentic Copilot tasks. The model is production-grade, the performance at scale is better, and the privacy guarantees are complete.
Use local when: you are working with code under HIPAA, GDPR, or financial compliance requirements; your environment is air-gapped or has strict network egress rules; your team runs enough agentic tasks that cloud token costs are measurable.
Stick with cloud when: you do not have a 24GB NVIDIA GPU; you are on Mac or AMD hardware; you want zero setup friction.
The full walkthrough is in Microsoft’s official Foundry Local post. The GitHub Copilot BYOK documentation covers how the model picker integration works across the CLI and VS Code extension.