A Foundry Local repository collaborator confirmed Qwen3.8-27B catalog availability on 4 November 2026. This is a new deployment option for an existing model: Qwen dates its release to 14 August.
For a development team, the useful question is whether that option fits its machines and approval rules. Moving model execution onto a workstation changes one part of the workflow; the coding client and its tools still need separate review.
Match the machine to the published setup
Microsoft’s 5 November guide specifies Foundry Local CLI 0.13.0 or later and NVIDIA hardware. Its 4-bit setup uses about 18 GB for the model; the guide specifies 24 GB VRAM for up to 64K context and 32 GB or more for up to 262K. Mac and AMD expansion is described as future work. These are vendor configuration figures, not our measurements.
The CLI 0.13.0 release, dated 2 November, is a preview. Its NVIDIA setup instructions require manual CUDA and compatible cuDNN installation, with different requirements for x64 and ARM64. Check the release’s architecture-specific instructions before changing an existing environment.
Do not turn the model’s memory figure into a purchasing guarantee. Our suggested first check is to record free VRAM while the editor, display workload and other normal applications are running. Then test representative repository tasks at the intended context size. A successful model load is only the beginning of an acceptance test.
Trace the agent’s connections
The Foundry Local FAQ says prompts and outputs at its endpoint are processed on-device. It separately identifies model/component downloads and optional diagnostic sharing as network uses.
GitHub’s BYOK documentation distinguishes user-configured model requests sent directly to the provider from enterprise-configured models routed through GitHub. Adding a custom model leaves other models available. Verify the actual client, configuration and selected model instead of assuming every Copilot workflow follows the same route.
A small pilot before a wider rollout
- Use a non-sensitive repository and record the model, runtime, driver and client versions. Keep a reproducible task set and expected results.
- Measure completed, correct changes alongside latency and peak memory. Run the project’s tests and review the diff; token throughput alone does not establish coding quality.
- List network-enabled tools, remote services and telemetry settings. Give the agent only the file and command permissions needed for the pilot, with human approval for consequential changes.
- Include hardware, energy and maintenance time in the cost comparison. A local inference endpoint does not make the entire development workflow cost-free or offline.
These are proposed checks, not hands-on results. We have not installed this preview, benchmarked the model or independently tested its privacy behavior.