Microsoft’s coding model MAI-Code-1.1-Flash can now run on your own PC inside GitHub Copilot, with no inference charge for local calls. The catch is the hardware: Microsoft recommends more than 120GB of RAM, so this is for high-end machines for now.
Here are the numbers that matter before you get excited.
- Local support in the Copilot app, Copilot CLI and VS Code is experimental, due by the end of October 2026.
- The on-device build is a 3-bit quantized model of about 53GB, with a 256K context window.
- Microsoft’s test machine was the Surface Laptop Ultra, with up to 128GB of unified memory.
- Local model calls carry no inference charge. Cloud calls are still billed as usual.
Short version: it’s real, it’s promising, and most laptops can’t run it yet.
What MAI-Code-1.1-Flash local mode actually does
Until now, Copilot sent your prompts to the cloud. With Microsoft’s October 7 announcement, Copilot can run the model on your own hardware instead. It does the same agentic coding work, such as reading files, editing code and running commands.
There’s also an Auto mode. Copilot picks between local and cloud inference for each task, and you can still pick a model by hand. If you want to see where this fits in Microsoft’s wider plan, our guide to Copilot Autopilot, Home and Code covers the cloud side.
One thing to know: “local” doesn’t mean “offline.” Model selection, inference and tool execution each have their own boundaries, so a session can mix local and cloud parts.
The hardware catch: why you need 120GB of RAM
Even compressed, the model is big. Microsoft says the quantized file is roughly 80% smaller than the full cloud version, yet it still weighs about 53GB. In Microsoft’s own test, memory use peaked at 75.5GB with a 256K context.
That’s why Windows Report calls it a model for workstations and AI-focused systems. This table shows how the two versions compare, based on Microsoft’s figures.
| Detail | Cloud version | Local (on-device) |
|---|---|---|
| Precision | BF16 | 3-bit quantized |
| SWE-Bench Verified | 72.6% | 70.80% |
| Terminal-Bench 2.1 | 62.9% | 66.29% |
| Context window | Cloud-hosted | 256K |
| Inference cost | Billed per use | No charge |
These are Microsoft’s own numbers, taken on a synthetic code-generation workload, so your results may differ. The local build scoring higher on one test and lower on the other is normal for this kind of benchmark.
The realistic target today is the Surface Laptop Ultra, which Microsoft lists with up to 128GB of unified memory and up to 1 petaflop of AI compute. Pre-orders are open, but it is a premium machine. Microsoft also describes a “hybrid intelligence” plan for Windows, where cloud and local models share the work. That suggests smaller local models will follow.
How to turn on local models and the sandbox
Microsoft says the Copilot app, Copilot CLI and VS Code all support local models. You can choose MAI-Code-1.1-Flash through the Windows ML provider, or point Copilot at any model on an OpenAI-compatible local endpoint.
The second option matters if your PC is smaller. A smaller model you already run locally can plug into the same setup, though the quality will depend on that model.
Microsoft also added a sandbox for agent sessions. In the Copilot app, open project settings and turn on “Sandbox new sessions.” In the CLI, run the /sandbox command.
The sandbox uses Microsoft Execution Containers, an open-source library. It needs no separate VM or container image. Shell commands and, by default, local MCP and language servers run inside it, while built-in file tools are checked by the agent harness instead.
Is local coding with MAI-Code-1.1-Flash worth it?
For privacy and cost, yes in theory. Code stays on your machine for inference, and heavy agent loops don’t rack up per-token bills. Microsoft reports prompt processing of about 923 tokens per second at 64k context.
The cloud model is also getting cheaper. Microsoft says MAI-Code-1.1-Flash costs a quarter of the version launched at Build in June. That narrows the savings from going local.
If you followed our post on GitHub Agentic Autofix, you’ll see the pattern: Microsoft and GitHub are pushing agents into every step of the coding workflow, and now onto your own hardware.
Frequently Asked Questions
Can I run MAI-Code-1.1-Flash on a normal laptop?
Probably not. Microsoft recommends more than 120GB of RAM for best performance, and its test peaked at 75.5GB. A typical 16GB or 32GB laptop falls far short.
Is local MAI-Code-1.1-Flash free?
Microsoft says local model calls have no inference charge. You still need a Copilot setup and the hardware, and cloud calls are billed as before. The announcement doesn’t say which Copilot plans get access.
When does local support arrive in GitHub Copilot?
Microsoft says experimental support in the Copilot app, CLI and VS Code is due by the end of October 2026. Its blog also says features are “rolling out now,” so check your own app.
Does local mode work fully offline?
Not necessarily. Inference can be local, but model selection and tool execution have separate boundaries, and Copilot’s Auto mode can still use the cloud.
Our take
Treat this as a preview of where AI coding is heading, not something to plan your week around. If you own a 128GB machine, try it. Everyone else should keep using cloud Copilot, watch for smaller local models, and use the new sandbox, which is useful on any PC.
Leave a Reply