GitHub Copilot to split coding work between local and cloud models on Windows

Auto orchestration and direct model choice

Microsoft says that, coming by the end of the month, GitHub Copilot will determine when a task is best handled by on-device intelligence and when it should use cloud-scale models. Developers do not have to manage infrastructure decisions themselves, as Copilot coordinates local and cloud inference behind the scenes.

GitHub is adding two ways to use local models across the GitHub Copilot CLI, the Copilot app and VS Code:

• Auto: Copilot chooses between local and cloud inference. Across a multi-turn session, it can consider task context and cache state while routing work, so useful cached work is preserved.
• Explicit selection: developers pick a specific provider, model or endpoint. They can select MAI Code 1.1 Flash through the Windows ML provider, or connect Copilot to OpenAI-compatible local endpoints and choose from the models those endpoints expose.

Microsoft stresses that local inference does not make a session offline, because model selection, inference and tool execution have different boundaries. The company describes the approach as the next step in the HydraFusion vision, an orchestration idea that spans multiple models and, now, multiple compute environments including the edge.

MAI Code 1.1 Flash on Surface Laptop Ultra

Microsoft AI developed a local version of MAI Code 1.1 Flash, a coding-optimized mixture-of-experts model with 137 billion total and 6.8 billion active parameters. The on-device work uses quantization and speculative decoding to shrink the footprint and improve responsiveness while keeping task completion and tool-use quality.

The hardware matters here. According to Microsoft, Surface Laptop Ultra is built around NVIDIA RTX Spark, with up to 128 GB of unified memory and up to 1 petaflop of AI compute. The authors note that unified memory is not all available to model weights, since the operating system, applications, the inference runtime and the key-value cache also need space. As an agent reads files and receives tool results, its context grows and memory use rises.

Microsoft reports that the quantized model comes in at 53GB, an 80% reduction in size compared with the Bfloat16 cloud variant. Peak memory usage reaches 75.5GB at 256k context, and prompt-processing throughput is 923.5 tokens per second at 64k context and 769.8 at 128k.

The company published benchmark results, tested October 5, 2026:

• SWE-Bench Verified (500 tasks): MAI Code 1.1 Flash 72.6%, GPT OSS 120B 32.0%, quantized on-device version 70.80%.
• Terminal-Bench 2.1 (89 tasks): MAI Code 1.1 Flash 62.9%, GPT OSS 120B 23.6%, quantized on-device version 66.29%.

The GPT OSS comparison used Unsloth's GPT-OSS-120B GGUF. The on-device test used mixed-precision quantization at approximately 3.3 bits per weight, DFlash2 sliding-window speculative decoding and a Windows ARM64 llama.cpp CUDA runtime. Microsoft notes that results may vary by device and configuration.

Sandboxing for agent tool execution

Microsoft points out that an agent's shell commands normally inherit the access of the account running them, and moving inference onto the device does not change that. Sandboxing applies a policy to the processes and local services the agent launches, controlling access to files, networks, credentials, system capabilities and execution paths, whichever model requested the work.

GitHub Copilot uses Microsoft Execution Containers (MXC), an open-source library from the Windows team that translates policy into native operating-system controls. On Windows it uses the BaseContainer tier of the ProcessContainer backend, on macOS Seatbelt, and on Linux bubblewrap. These backends need no separate virtual machine or container image.

With sandboxing enabled, shell commands and, by default, local Model Context Protocol servers and language servers run inside the process boundary. Built-in file tools run inside Copilot itself, and their requests are checked against the effective policy, but Microsoft says this is not OS-enforced child-process isolation. Remote MCP servers sit outside the local process sandbox. Users can configure settings at any time by running the `/sandbox` slash command in the Copilot CLI.

What it means for developers

Microsoft frames the update around choice and control: developers can let Copilot place work automatically or pin a specific local model, while sandbox policies set clear limits on what agents can touch. The company gives the example of a daily job that reads local repositories, runs tests in working copies and writes an HTML report, with source repositories kept read-only and no network access for test processes. The same job can use either a local or a cloud model.

The local experience is described for NVIDIA RTX Spark Windows PCs such as Surface Laptop Ultra. The source does not give pricing or further hardware availability.