GitHub Copilot gained local AI coding on new Windows PCs

GitHub Copilot will soon decide whether a coding task should run on the device or in the cloud, and Microsoft says that local mode is now arriving on new Windows PCs.

The setup targets NVIDIA RTX Spark Windows PCs such as Surface Laptop Ultra. On those machines, Copilot can use local inference models instead of sending every step away to a server.

Microsoft says the system can also choose between local and cloud models on its own. Developers can still override that choice and pick a local model directly when they want tighter control.

The local model is MAI Code 1.1 Flash, a coding model with 137 billion total parameters and 6.8 billion active parameters. Microsoft says the on-device version uses quantization and speculative decoding to cut its footprint and speed up responses.

Memory is the big constraint here. On Surface Laptop Ultra, Microsoft says peak memory use reached 75.5GB at 256k context, and prompt processing reached 923.5 tokens per second at 64k context and 769.8 tokens per second at 128k.

The company also says the quantized on-device version is 53GB, which is an 80% reduction compared with the Bfloat16 cloud version. In plain English, that is the difference between fitting in the backpack and needing a moving truck.

Copilot is also getting sandboxed tool execution through Microsoft Execution Containers, or MXC. That means shell commands and some local services can run with tighter limits on files, networks, credentials, and system access.

Microsoft says developers can use the GitHub Copilot CLI, the Copilot app, and VS Code with two paths: let Auto route tasks, or select a local model directly. It also says OpenAI-compatible local endpoints can be connected for models exposed there.

The example Microsoft gives is a daily repository dashboard that reads local repos, runs tests, and writes an HTML report while keeping source repos read-only and blocking network access for tests. The source does not show how broadly these features are rolling out beyond the named Windows PCs, and it does not test every coding workload.