Copilot on Windows will hand some tasks to local models
Microsoft said on October 7 that Copilot on Windows will be able to read your files and recent activity, organize files, run device diagnostics and hand some tasks to models running on the PC itself. It all needs a Copilot+ PC and your permission, and the rollout is only promised “in the coming months”. The first piece developers can try is local-model routing in GitHub HydraFusion, which comes to the GitHub Copilot app, GitHub Copilot CLI and Visual Studio Code as an experimental preview later in October.
Microsoft calls the plan “hybrid intelligence”: Windows runs each task either on the device or in the cloud, whichever fits. The company’s reasoning, from Microsoft’s Windows Experience Blog, is about money as much as capability: “customers’ needs are outpacing what their cloud budgets can support.” For Copilot, the promise is “more capability while helping your tokens go further.”
What Copilot will do, and on which PCs
Microsoft splits the Copilot upgrade into three local capabilities. Local context lets Copilot, with your permission, see relevant content on the PC, including files and recent activity. Local actions cover tasks across Windows: organizing files, assessing device diagnostics, troubleshooting, coding and running workflows on the device. Local models let Copilot use on-device compute for some tasks and pull in the cloud when it needs to.
These apply across Home, Code and Autopilot, the three modes of the new Copilot, where Code runs on the same tech as GitHub Copilot and Autopilot is Scout Agent, Microsoft’s proactive assistant, under a new name. In Home, Copilot pulls in the files you have been working on and turns them into “collaboration-ready artifacts”. In Code, it builds native Windows apps from a single prompt, runs code inside Microsoft Execution Containers and uses local models “when appropriate” to keep token costs down. Autopilot, which Microsoft describes as “a persistent, proactive and personal agent”, gets local context, local models and PC actions so it can keep working while you do something else.
Copilot+ is a hardware class. Microsoft’s own definition includes an NPU, a chip dedicated to AI workloads, capable of over 40 trillion operations per second, according to a 2024 Windows Experience Blog post. Microsoft now says over 40% of laptops being built for business are Copilot+ PCs, and that those machines run more than 2 trillion inferences a month locally, mostly for search, imaging and video calls.
What arrives this month
GitHub launched HydraFusion earlier this year to route each task to the right model in the cloud. On Windows it will also be able to send tasks to models running on the device, and that is the part landing in experimental preview later in October. Testing Catalog reports that Microsoft’s companion post on GitHub Copilot covers the local memory requirements of MAI Code 1.1 Flash and offers explicit selection of a local model as well as automatic routing.
Microsoft Execution Containers (MXC) is already generally available on Windows 11. It is Microsoft’s sandbox for agents: organizations define which files and networks an agent can reach, and the policy is enforced at runtime. Isolation ranges from process and session isolation to WSLc, virtual machines and Windows 365 for Agents. MXC also works on other operating systems, but Microsoft says the Windows integration gives organizations more control and more containment options. Agent identity lets IT tell an agent’s actions apart from the user’s, and that activity feeds into Agent 365 and Intune for policy and monitoring.
Codex from OpenAI, GitHub Copilot, OpenClaw, Replit, LM Studio, OpenShell from NVIDIA and Unsloth AI support MXC today. Claude Code, Perplexity, Raycast, Manus, Box and Hermes Agent by Nous Research are among those listed as coming, and Meta’s personal agent Muse for Windows is due as a native app with MXC built in. For OpenClaw on mini desktop PCs, Microsoft is adding a native Windows gateway that replaces terminal-based setup and runs the agent in an MXC sandbox.
Windows ML, Microsoft’s runtime for deploying models across GPU, NPU and CPU, is getting llama.cpp support, which opens it to the open-source models that ship in that format.
How big “local” got
MAI Code 1.1 Flash, introduced at Build, has 137 billion total and 6.8 billion active parameters. Stored at 3-bit precision it shrinks by nearly 80%, and Microsoft says coding quality holds, with a 256K context window running locally. For scale, Phi Silica, the on-device model Microsoft built for Copilot+ PCs in 2024, has 3.3 billion parameters and a 4k context length, so MAI Code’s total parameter count is more than 40 times larger.
Other models on Microsoft’s list target NVIDIA’s RTX Spark hardware: an upcoming Nemotron model from NVIDIA with over 70 billion parameters, quantized to 2-bit so it needs just over 20GB of memory, and DeepSeek V4 Flash at 284 billion parameters. Phi Silica itself is on its way out. Microsoft’s documentation says a new on-device model, Aion Instruct, replaces it, reaching Windows Insider devices in November and retail devices in January 2027, when Phi Silica is removed.
Surface Laptop Ultra, powered by RTX Spark, went up for pre-order on Wednesday alongside new machines from ASUS, Dell, HP, Lenovo and MSI. Testing Catalog reports it takes up to 128 GB of unified memory, runs models above 120 billion parameters and arrives October 16, with DGX Station for Windows systems, rated for models up to one trillion parameters, due later this year.
Up next
Claude Max and Team get API credits that skip Claude Code, Video
Max plans picked up monthly API credits, $100 on Max 5x and $200 on Max 20x, and none of it works in interactive Claude Code sessions. It’s reserved for the apps and agents you build on the Claude API.
Related stories
- Microsoft Copilot gets a Code mode on GitHub Copilot tech
- Surface Laptop Ultra slows down on local AI once unplugged
- Maia 300 promises Microsoft 40% more tokens per megawatt
- Claude Code on desktop folds an iPhone Duo to test layouts
- Hugging Face sends language model calls via OpenRouter
- Cursor hosts and deploys apps its cloud agents build
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

