NVIDIA and Microsoft Pair Windows AI PCs With New Agent Controls

Microsoft Adds Local Coding to GitHub Copilot, With Cloud Routing Due by October’s End

Microsoft is bringing coding AI onto developers’ PCs, but that does not make an entire Copilot session offline or confined to the device. In its October 7 announcement, the company said local models and sandboxed tools are rolling out now. Automatic routing between local and cloud models is promised by the end of October 2026.

The change extends GitHub’s HydraFusion approach beyond choosing among models to choosing where they run. Microsoft says Copilot can consider the task, conversation context and cached work as it moves between local and cloud inference—the process of running a model to produce an answer.

Across the Copilot command-line interface, Copilot app and VS Code, Microsoft is adding two ways to use local models. They serve different needs rather than requiring everyone to hand placement decisions to the router:

Auto lets Copilot choose local or cloud processing throughout a multistep conversation, taking account of useful cached work.

Explicit selection lets developers choose MAI Code 1.1 Flash through Windows ML, or connect to an OpenAI-compatible local endpoint and select a model it exposes.

Microsoft’s on-device MAI Code 1.1 Flash has 137 billion total parameters, with 6.8 billion active. Its local build uses quantization, which reduces numerical precision to shrink memory requirements. Speculative decoding adds a drafting step: one model proposes blocks of output that the target model checks.

Microsoft says the compressed build occupies 53 GB, an 80% reduction from its Bfloat16 cloud version. The first shipping version runs on Surface Laptop Ultra, built around NVIDIA RTX Spark with up to 128 GB of unified memory shared by the CPU and GPU.

That footprint is not the whole memory bill. The operating system, applications and runtime also need space, as does the cache holding information from earlier tokens. Microsoft reports peak memory use of 75.5 GB with a 256,000-token context. Keeping the model loaded avoids repeated loading, but does not guarantee steady response times.

A local model does not remove an agent’s ability to issue shell commands with the user account’s permissions. Copilot uses Microsoft Execution Containers, an open-source library that applies access rules through native operating-system controls. Microsoft’s Windows Developer Blog says the library is now generally available and keeps its policy outside the agent workload’s control.

When sandboxing is enabled, shell commands and, by default, local Model Context Protocol servers and language servers run inside the process boundary. Model Context Protocol connects agents to tools and services. Built-in file tools instead receive policy checks inside Copilot; they are not isolated as operating-system-enforced child processes.

Remote tool servers remain outside that local process sandbox, with connection-policy checks handled inside Copilot when applicable. Developers can configure sandbox settings using the /sandbox command in Copilot’s command-line interface, separately from choosing which model handles the work.