# Foundry Local Runs Qwen3.8-27B in Copilot: Hardware and Preview Limits Microsoft's local coding guide connects Qwen3.8-27B to GitHub Copilot. Check NVIDIA memory, preview runtime requirements and what its performance claims establish. By Paras Madan | 2026-11-05 Microsoft has published a [Foundry Local coding guide](https://devblogs.microsoft.com/foundry/foundry-local-agentic-coding-qwen3-8-27b/) for running Qwen3.8-27B on NVIDIA hardware and connecting it to GitHub Copilot. The November 5 announcement recommends at least 24 GB of VRAM and Foundry Local CLI 0.13.0 or later. Microsoft puts the four-bit model at roughly 18 GB. For developers with a suitable GPU, this provides a documented route from a local model server into an existing coding interface. Buying hardware for it deserves a more careful evaluation than installing it on a machine you already own. ## The CLI is still a preview The [CLI 0.13.0 release](https://github.com/microsoft/foundry-local/releases/tag/cli-preview-0.13.0) was published on November 2 at 22:18 UTC. Today's announcement explains the coding setup built on that earlier release. Qwen3.8-27B is also an existing model. Microsoft labels the CLI build a public preview and warns that behavior can change. It includes Foundry Local SDK and Runtime 2.3.0, ONNX Runtime 1.31.0 and ONNX Runtime GenAI 0.18.0, along with coding-model support and fixes for shutdown and memory errors. NVIDIA dependencies need separate installation: - Windows and Linux x64 require CUDA 12 and compatible cuDNN 9 runtimes. - Windows and Linux ARM64 require CUDA 13 and compatible cuDNN 9 runtimes. - Runtime library locations must be configured, followed by a server restart. The CLI has a macOS package, but Microsoft's Qwen3.8 coding announcement describes Mac and AMD GPU support as future work. An installer being available for an operating system does not establish that this particular accelerated configuration works there. ## Local requests through a familiar client [GitHub's BYOK documentation](https://docs.github.com/en/copilot/concepts/models/bring-your-own-key) says user-configured models receive requests directly from the client rather than through GitHub's Copilot API. Supported clients can use local or cloud providers, and this arrangement does not require a Copilot subscription. Availability depends on the client and configuration method. An offline setup still needs a reachable model endpoint. Repository hosting, package downloads and external tools are separate dependencies, so test the whole development workflow before relying on it without a connection. The practical privacy check is similarly specific: confirm which provider is selected and where its endpoint runs. Review any other tools the coding agent can call before giving it confidential material. ## Speed claims need a task-level check Microsoft's optimizations include DFlash2 speculative decoding, prefix caching and continuous batching. Its tests report faster generation than llama.cpp on an RTX 5090 and DGX Spark, with a larger advantage beyond 131K prompt tokens. These are vendor results for particular configurations, not an independent comparison of completed coding tasks. The [Qwen model card](https://huggingface.co/Qwen/Qwen3.8-27B) lists Apache 2.0 licensing and a native 262,144-token context window. Its benchmark notes describe harness and evaluation settings, including Claude Code for several coding tests. Those scores do not measure this quantized Foundry Local and Copilot combination. To decide whether to use it, choose a few issues with known solutions. Measure time until the tests pass, inspect the resulting diff and record peak memory. Include a small change and a task that requires reading several files. Repeat one task with a warm server to see whether cached context helps your actual work. That exercise is more useful than comparing output speed alone. A faster stream of tokens can still require extra corrections before a change is ready to merge. I'm Paras. I work at the intersection of AI, engineering and distribution. I help technical founders turn AI infrastructure and developer-tool updates into clear technical writing and product stories. [Book a 30-minute call](https://cal.com/paras-madan/30min).