GitHub Copilot on Windows gets local models, with HydraFusion routing between device and cloud

Microsoft and GitHub announced on October 7, 2026 that the GitHub Copilot coding assistant on Windows can use local models, meaning models that run on the PC itself. At the center of the announcement is GitHub's HydraFusion, a routing layer added to Copilot earlier this year that sends each task to the most suitable cloud model. According to Microsoft's announcement, it will now also be able to use models running on the device.

The goal is to handle part of the coding work on the PC without sending it over the internet, and to save cloud usage for the tasks that really need it. Microsoft calls this approach "hybrid intelligence": agents run locally when it makes sense and in the cloud when needed. The feature is coming to the GitHub Copilot app, Copilot CLI and Visual Studio Code.

Two ways to use local models

According to the technical post on Microsoft's Command Line blog, local models can be used in two ways. With automatic routing, Copilot decides for each task whether to run it on the device or in the cloud; this is expected by the end of October. With explicit selection, developers pick a specific local model themselves: MAI Code 1.1 Flash through Windows ML, or any OpenAI-compatible local endpoint. The second option lets developers who run their own server, for example with llama.cpp, connect their models to Copilot. The post also notes that local inference does not make the session fully offline.

For each request, HydraFusion chooses one of three patterns: give the task to a single model; have a cheaper, faster model write a draft and escalate to a stronger model if it fails a quality check; or have a model from a different family review a first model's draft, which is then revised. According to GitHub's community page, HydraFusion only uses models that the user's plan and organization allow; it has no separate fee, and each model it calls is billed at its standard rate.

Models that run on the device

The main local model Microsoft highlights is MAI Code 1.1 Flash, its coding model. According to the company, the model has 137 billion total and 6.8 billion active parameters; its on-device version is compressed to about 3.3 bits per weight, bringing it down to 53 GB, and keeps a 256K-token context locally. In Microsoft's comparison, the compressed local version scores close to the cloud version on SWE-Bench Verified and slightly higher on Terminal-Bench 2.1. The announcement also names an upcoming NVIDIA Nemotron model and DeepSeek V4 Flash among the local options.

Models of this size need powerful hardware. Microsoft's example device is the Surface Laptop Ultra with NVIDIA RTX Spark and up to 128 GB of unified memory; according to the company, MAI Code 1.1 Flash peaks at 75.5 GB of memory on it at a 256K context. RTX Spark PCs from ASUS, Dell, HP, Lenovo, MSI and Surface go on sale on October 16. llama.cpp support is also being added to Windows ML.

Agents run in an isolated environment

Letting local agents run commands on a PC raises security questions. For this, Microsoft uses Microsoft Execution Containers (MXC), now generally available on Windows 11. MXC is an open-source library that turns rules about which files and networks an agent may access into the operating system's own controls. When sandboxing is on in Copilot, shell commands and, by default, local MCP servers run inside that boundary. It can be turned on with the /sandbox command in the CLI or the "Sandbox new sessions" option in project settings in the Copilot app. According to Microsoft, agents supporting MXC include OpenAI's Codex, GitHub Copilot, OpenClaw, Replit, LM Studio and Unsloth AI, with Anthropic's Claude Code and Perplexity adding support.

Automatic routing is expected to arrive as an experimental preview, and, as Microsoft notes, local model performance can vary by device and configuration. The published test results come from Microsoft, and large local models run comfortably only on new PCs with a lot of memory. It is worth testing how well a local model handles your own codebase. For a fuller picture, follow evaluations from independent researchers as well.

Sources: Windows Experience Blog: Building Windows for hybrid intelligence, Microsoft Command Line: Bringing local models and sandboxed tools to Windows and GitHub Copilot, GitHub Community: HydraFusion.