Windows hybrid intelligence is Microsoft’s plan to let AI features run on your PC when they can and reach the cloud only when they need to. Microsoft laid it out at its October 7 Windows and Surface event, and the first piece for developers arrives later this month.
In short: your PC picks between a local model and a cloud model for each task. Most of it targets Copilot+ PCs and developers for now, so don’t expect your current laptop to change overnight.
Here are the quick facts:
- Developers get HydraFusion-based routing in an experimental preview “later this month” (October 2026).
- Copilot hybrid features come to Copilot+ PCs “in the coming months.”
- Copilot can only use your local files and activity if you give permission.
- Windows ML is adding llama.cpp support for open-source models.
- RTX Spark PCs, including the Surface Laptop Ultra, ship from October 16.
That list is the whole announcement in miniature. The rest of this post explains what each piece means for you.
What Windows hybrid intelligence actually does
Microsoft describes it as agents that “can run locally when it makes sense, reach the cloud when they need to.” Think of it as a traffic cop for AI requests. Small, private jobs stay on your machine. Hard ones go to a bigger cloud model.
Why bother? Cloud AI costs money and sends your data off the device. Running some work locally cuts both, and it can keep working with a weak connection.
On Copilot+ PCs, Microsoft says Copilot gets three new abilities, which are worth seeing side by side.
| Ability | What it means | What you control |
|---|---|---|
| Local context | Copilot can read your files and recent activity on the PC | Needs your permission |
| Local actions | Organizing files, diagnostics, troubleshooting, coding | Microsoft says you stay in control at every step |
| Local models | Small models on the device, combined with cloud models | Routing is automatic; no off-switch details yet |
The table shows the bargain: more help, but only with your say-so. Microsoft’s own post says Copilot’s search integration is opt-in too.
HydraFusion: the router behind the idea
GitHub launched HydraFusion earlier this year to send each coding task to the best cloud model. Microsoft is now extending it so it can also pick models running on your device.
Per Microsoft’s announcement, the experimental preview lands in the GitHub Copilot app, GitHub Copilot CLI, and Visual Studio Code later in October. If you code with Copilot, that’s where you’ll see it first.
Microsoft hasn’t said how HydraFusion decides between local and cloud. Until it does, treat the routing as a black box and watch the preview notes.
Which local AI models are in the mix
Microsoft named three models, and each one shows a different trick for fitting big AI into a PC.
- MAI Code 1.1 Flash runs at 3-bit precision, cutting its size by nearly 80% while keeping a 256K context window.
- An upcoming NVIDIA Nemotron model has over 70 billion parameters, squeezed to 2-bit so it needs just over 20GB of memory.
- DeepSeek V4 Flash, at 284 billion parameters, was shown for RTX Spark hardware.
The takeaway is memory. Quantization shrinks models, but you still need plenty of RAM, which is why the high-end RTX Spark machines lead this push. We covered the hardware in our RTX Spark laptop price list.
What changes for developers with Windows ML
Windows ML is Microsoft’s runtime for running models across GPUs, NPUs, and CPUs. It’s now gaining llama.cpp support.
That matters because llama.cpp is the go-to engine for open-source models. Developers can try new releases on Windows soon after they ship, without waiting for a custom port.
Privacy and safety: what Microsoft says it will do
Local inference can keep sensitive data inside a company’s own environment. Microsoft also says agent activity will be distinguishable from yours, and IT teams can set policy through Agent 365 and Intune.
Microsoft Execution Containers are also now generally available on Windows 11, letting organizations limit which files and networks an agent can touch. Our guide to Microsoft Execution Containers covers how that works.
One gap: Microsoft’s post doesn’t spell out a single switch to turn hybrid features off. Check Settings when the Copilot rollout reaches you.
Frequently Asked Questions
Is Windows hybrid intelligence the same as Windows 12?
No. Windows 12 did not debut at the event, and the hybrid features are updates to Windows 11.
Do I need a new PC for local AI on Windows?
For the Copilot features, Microsoft targets Copilot+ PCs. Bigger local models need lots of memory, like the RTX Spark machines with up to 128GB.
When can I try HydraFusion on Windows?
Developers can expect an experimental preview in the GitHub Copilot app, Copilot CLI, and VS Code later in October 2026.
Can Copilot read my files without asking?
Microsoft says access to local context depends on your permission, and Copilot actions keep you in control at every step.
Our take
This is a sensible direction, but it’s early. If you’re a developer, try the preview and see whether local routing really saves you time. If you’re a regular user, wait for the Copilot rollout and read the permission prompts before saying yes. And for a look at another Windows AI feature in testing, see our post on Windows 11 Search actions.
Leave a Reply