Image: Microsoft

Microsoft wants Copilot on Windows to do more of its work on your own PC. In a Windows blog post, Windows chief Pavan Davuluri says Copilot will be able to read relevant files and recent activity on your computer, take actions across Windows, and run some tasks on AI models stored on the device instead of in the cloud. Microsoft calls this "hybrid intelligence."

We’re supercharging Copilot on Windows with Hybrid Intelligence.

With your permission, Copilot can tap into the context on your PC, take action for you, and use local models when it makes sense, giving you more capability while helping your tokens go further. pic.twitter.com/7YvqHh8RKT

— Satya Nadella (@satyanadella) October 7, 2026

Microsoft's examples include organizing files, checking device diagnostics, troubleshooting problems and turning the files you've been working on into shareable documents. Autopilot, the always-on agent Microsoft introduced last month, would also get local context and the ability to act on the PC.

Not every Windows user will get this. The features are limited to Copilot+ PCs, which need an NPU rated at 40+ TOPS, and Microsoft only says they're "expected to begin rolling out" in the coming months, with timing varying by device, market and chip.

Several important details are missing. Microsoft doesn't say which local models Copilot will use, how it decides between local and cloud, whether your files are sent to cloud models when a task goes there, or what is stored. It describes the access only as "with your permission."

That matters because Microsoft has been here before. In 2024, it reworked Recall, a Copilot+ feature that saves snapshots of your screen, making it opt-in and off by default and delaying its launch after privacy and security concerns.

Nadella's "tokens" line is about money. Microsoft moved Copilot's Cowork, Code and Autopilot to usage-based billing last month. Running some of that work on your own chip could cut what businesses pay, but Microsoft hasn't said how local runs will be counted or priced.

For businesses, the more concrete news is security. Microsoft Execution Containers (MXC) are now generally available on Windows 11. They let organizations set which files and networks an AI agent can reach, enforced while it runs. Microsoft says OpenAI's Codex, GitHub Copilot, OpenClaw and LM Studio already support MXC, while Anthropic's Claude Code, Perplexity, Manus and Hermes Agent are among those promising support.

Developers get local routing too. GitHub's HydraFusion, which picks a cloud model for each task, will be able to use on-device models in the GitHub Copilot app, Copilot CLI and VS Code as an experimental preview later this month.

Bigger local models still need serious hardware. Microsoft is shrinking its MAI Code 1.1 Flash coding model, which has 137 billion parameters, to 3-bit precision. By our math, that's still about 51GB of weights, more than double the memory in the base Surface Laptop Ultra.

Microsoft says the Laptop Ultra can run models with more than 120 billion parameters, but that relies on configurations with up to 128GB of memory. The 128GB model costs $5,899.99, according to The Verge, which also lists the base model at $2,599 with 24GB.

The Laptop Ultra and NVIDIA RTX Spark laptops from ASUS, Dell, HP, Lenovo and MSI are on pre-order now and ship later this month. Microsoft and NVIDIA claim up to 6.2x faster AI video generation than a MacBook Pro with M5 Pro, based on their own tests of preproduction machines.

Until the Copilot features ship, none of this has been independently tested. Microsoft says it will share more at Ignite in November.