Yes, you can now run a serious coding model on your own PC and plug it into GitHub Copilot. Microsoft just showed how to do it with Qwen3.8-27B Foundry Local, and the catch is that you need an NVIDIA GPU with at least 24 GB of VRAM.
Here are the facts you need before you spend an evening on it.
- Announced November 5, 2026, on the Microsoft Foundry blog
- Needs an NVIDIA GPU with 24 GB VRAM or more
- Model size is about 18 GB at 4-bit quantization
- Requires Foundry Local CLI 0.13.0 or later
- Documented for Windows only, for now
If you only skim one section, make it the GPU table below. That is where most people will find out whether this is for them.
What Qwen3.8-27B Foundry Local actually gives you
Foundry Local is Microsoft’s tool for running AI models on your own machine. The new guide, published on the Microsoft Foundry blog, points it at Alibaba’s Qwen3.8-27B and then connects that model to GitHub Copilot.
The connection uses Copilot’s Bring Your Own Key feature. Instead of sending your code to a cloud model, Copilot talks to a local endpoint on your computer.
Microsoft lists three reasons to bother: predictable inference costs, more privacy, and the ability to work offline. If you have ever hit a usage cap mid-task, the first one will sound familiar.
This is the same idea we covered in Android Studio Rabbit 2 and its bring-your-own-agent option: your editor stays the same, and you choose what powers it.
Which GPU do you need for local Copilot?
The model file is about 18 GB, but you also need room for the context window. More VRAM means Copilot can remember more of your project at once.
| Your VRAM | Max context | What it means |
|---|---|---|
| 24 GB | Up to 64K tokens | Enough for small and mid-size tasks. The minimum for this setup. |
| 32 GB or more | Up to 262K tokens | The model’s full context length, so it can hold far more code. |
| Under 24 GB | Not covered | Microsoft does not document this setup for smaller cards. |
Microsoft also says you must install compatible CUDA and cuDNN runtimes yourself. Foundry Local does not download them, and on Windows you need to add their DLL folders to your PATH.
Memory is tight across the PC market right now, so this hardware is not cheap. Our laptop RAM shortage guide explains why prices are where they are.
How to set it up in four steps
The steps below follow Microsoft’s walkthrough, which uses the GitHub Copilot app on Windows.
- Install the CLI with
winget install Microsoft.FoundryLocal, then check it withfoundry --version. You need 0.13.0 or later. - Load the model with
foundry model load qwen3.8-27b. It downloads the first time. - Run
foundry server statusand copy the endpoint it shows. Use your own port, not the one in Microsoft’s screenshot. - In the Copilot app, open Settings, then Model providers, then Add provider, and choose Foundry Local. Paste the endpoint as the Base URL and add Qwen3.8-27B from the catalog.
Finally, pick Foundry Local as the provider and Qwen3.8-27B in the model picker. If you changed PATH while Foundry Local was running, run foundry server restart first.
How good is it, and how fast?
Microsoft quotes benchmark scores that Qwen published for the model. They are the vendor’s own numbers, so treat them as a rough guide rather than a verdict.
| Benchmark | What it tests | Score |
|---|---|---|
| Terminal Bench 2.1 | Agentic terminal coding | 73.0 |
| SWE-bench Pro | Agentic coding on real issues | 61.7 |
| NL2Repo-Bench | Repo-level code generation | 42.3 |
| DeepSWE 1.1 | Agentic coding | 42.2 |
On speed, Microsoft says Foundry Local generated text faster than llama.cpp at every prompt length it tested. The gap grew past 131K tokens.
The tests ran on an RTX 5090 and an NVIDIA DGX Spark, using real GitHub issues. The post only shows exact tokens per second in charts, so we cannot quote a number.
Microsoft is also upfront about the trade-off. A 27B model reasons less well than the biggest cloud models, so the authors suggest a hybrid approach: local for everyday work, cloud for the hard problems. For another free agent option, see our look at DeepSeek Super-agent.
Frequently Asked Questions
Can I run Qwen3.8-27B on a Mac or an AMD GPU?
Not yet from this guide. Microsoft says it plans to expand Qwen3.8-27B support to Macs and AMD GPUs, but gives no date.
Does the local model work offline?
Yes, once the model is downloaded. Microsoft lists offline use as one of the main benefits.
Does it cost anything?
The post does not state pricing or licensing. You still need the hardware, and you may still need a Copilot account to use the app.
Is it as smart as cloud models?
No. Microsoft says local models of this size have weaker reasoning than top cloud models, and recommends mixing the two.
Should you try it?
If you already own a 24 GB or 32 GB NVIDIA card and care about privacy or usage limits, try it this weekend. The setup is four commands and one settings screen. If you would have to buy a GPU just for this, wait: the Mac and AMD support Microsoft promised could change the math, and a cloud plan is cheaper until then.
Leave a Reply