Skip to content

NewsGPU

Microsoft Lets GitHub Copilot Use a Local Coding Model on Your Own GPU via Foundry Local

Microsoft says GitHub Copilot can now use Qwen3.8-27B running locally through Foundry Local on NVIDIA graphics cards. At 4-bit the model is about 18GB, so Microsoft recommends 24GB of VRAM for 64K tokens of context and 32GB or more for longer contexts. Speed figures are Microsoft's own. Forge PC has not tested this setup.

By , Editorial team

3 min read

Microsoft Lets GitHub Copilot Use a Local Coding Model on Your Own GPU via Foundry Local

Microsoft has shown how to plug a locally run AI model into GitHub Copilot, so your code never has to leave your PC. In a Foundry blog post dated 5 November 2026, Microsoft says its Foundry Local runtime can now run Qwen3.8-27B, a 27-billion-parameter open coding model, and serve it to GitHub Copilot through Copilot's Bring Your Own Key (BYOK) feature.

Forge PC has not tested this setup. The details below come from Microsoft's Foundry blog and GitHub's documentation, with UK graphics card prices from SpecPrice checked on 15 November 2026.

What Microsoft announced

According to Microsoft, Foundry Local optimises the whole Qwen3.8-27B pipeline on NVIDIA GPUs. The post lists several techniques, including speculative decoding with a small "drafter" model, 4-bit (INT4) weights, a quantised KV cache and CPU offloading.

The key numbers from the blog:

  • With 4-bit quantisation the model is about 18GB, which Microsoft says makes it runnable on consumer GPUs with 24GB of VRAM, with context windows of up to 64K tokens.

  • GPUs with 32GB or more can use longer contexts, up to the model's maximum of 262K tokens.

  • You need Foundry Local CLI 0.13.0 or later, plus NVIDIA's CUDA and cuDNN runtimes, which Foundry Local does not install for you.

Microsoft also compared Foundry Local with llama.cpp running the same 4-bit model on an RTX 5090 and on an NVIDIA DGX Spark, using real GitHub issues. It says Foundry Local delivered higher generation speed at every prompt length it tested, with the gap widening beyond 131K tokens. That is Microsoft's own test, not an independent result.

How it works with GitHub Copilot

Microsoft's walkthrough uses the GitHub Copilot app:

  1. Install Foundry Local with winget install Microsoft.FoundryLocal and check the version.

  2. Load the model with foundry model load qwen3.8-27b. This downloads it the first time and returns a local API endpoint.

  3. In the GitHub Copilot app, open Settings, then Model providers, choose Add provider and select Foundry Local.

  4. Paste the endpoint into the Base URL field, find Qwen3.8-27B in the catalogue and add it.

  5. Switch the model picker to Foundry Local and Qwen3.8-27B.

Microsoft is upfront about the limits. It says local models "may have limited reasoning capabilities compared with advanced cloud models due to their smaller size", and suggests using cloud models for complex tasks and the local model for everyday coding.

The hardware catch for UK buyers

The 24GB VRAM requirement is the real story for most readers. Microsoft's instructions are for NVIDIA GPUs only for now; the company says it plans to add Macs and AMD GPUs later.

In the UK, 24GB and 32GB cards are scarce and expensive. On 15 November 2026, SpecPrice found no stock-checked new listing for any 24GB desktop card at the UK shops it tracks, and its cheapest RTX 3090 was about £2,140 at PcNova, from a feed it could not stock-check. The cheapest in-stock RTX 5090 (32GB) was about £7,000 at Scan. Our guide to the best GPUs for local AI and our advice on buying a used GPU for local AI explain the cheaper routes, and why VRAM matters more than almost anything else.

If you have a smaller card, there are other ways to get a local coding assistant. Qwen3.8-27B is also listed in Ollama's library at 18GB, as we noted in our vision models guide, and you can pair Ollama with VS Code using Continue or the terminal tool Aider, with smaller models that fit your VRAM.

Our verdict

Foundry Local's GitHub Copilot integration is a genuinely useful option if you already own an NVIDIA card with 24GB or more: Microsoft says the 4-bit Qwen3.8-27B model is about 18GB and supports up to 64K tokens of context on 24GB cards. Its speed claims are Microsoft's own, and the company admits local models trail cloud models on harder reasoning. For most UK buyers, the 24GB requirement is the stumbling block, because those cards cost thousands of pounds on 15 November 2026. Forge PC has not tested this setup. New to Foundry Local? Start with our Foundry Local setup guide.

The bottom line

Foundry Local's GitHub Copilot integration is a genuinely useful option if you already own an NVIDIA card with 24GB or more: Microsoft says the 4-bit Qwen3.8-27B model is about 18GB and supports up to 64K tokens of context on 24GB cards. Its speed claims are Microsoft's own, and the company admits local models trail cloud models on harder reasoning. For most UK buyers, the 24GB requirement is the stumbling block, because those cards cost thousands of pounds on 15 November 2026. Forge PC has not tested this setup. New to Foundry Local? Start with our Foundry Local setup guide.

Sources

  1. 01Microsoft Foundry Blog, Bring Agentic Coding On-Device with Foundry Local (5 November 2026)
  2. 02SpecPrice, Cheapest RTX 3090 UK price
  3. 03SpecPrice, Cheapest RTX 5090 UK price
  4. 04SpecPrice, Cheapest 24GB GPU in the UK
  5. 05Ollama, qwen3.8 model library page

About the author

Forge PC Staff

Editorial team

The shared byline for articles researched, written and edited by the Forge PC editorial team. How we source, check and correct our work is set out in our editorial policy.

More from Forge PC Staff

Read next

How to Keep Ollama Models Loaded (or Unload Them to Free VRAM)

How-To

How to Keep Ollama Models Loaded (or Unload Them to Free VRAM)

Ollama unloads models five minutes after your last request. Use ollama ps to see what's loaded, ollama stop to free VRAM straight away, the keep_alive API parameter per request, and OLLAMA_KEEP_ALIVE or OLLAMA_MAX_LOADED_MODELS to change the defaults.

Forge PC Staff3 min

RTX 5070 vs RTX 5070 Ti for Local AI: Is 16GB Worth the Step Up?

Versus

RTX 5070 vs RTX 5070 Ti for Local AI: Is 16GB Worth the Step Up?

For local AI, the RTX 5070 Ti's 16GB of GDDR7 on a 256-bit bus matters more than its extra CUDA cores. NVIDIA lists the RTX 5070 with 12GB on a 192-bit bus, 988 AI TOPS and 250W, against 1,406 AI TOPS and 300W for the Ti.

Forge PC Staff2 min

How to Work Out What Running Local AI Costs in Electricity (UK, 2026)

How-To

How to Work Out What Running Local AI Costs in Electricity (UK, 2026)

Multiply watts by hours, divide by 1,000 and multiply by your unit rate. At Ofgem's average cap rate of 26.32p per kWh (1 October to 31 December 2026), an RTX 5090 at its full 575W rating costs about 15p an hour. Measure your own PC with nvidia-smi or a plug-in energy monitor.

Forge PC Staff4 min