Microsoft puts hybrid intelligence at the heart of Windows

Microsoft is pitching hybrid intelligence as a combination of local models, a routing mechanism, an AI runtime and silicon, so that work a PC can handle stays on the PC and heavier tasks go to the cloud. The company says the aim is to reduce tokens consumed in the cloud while keeping the AI performance users need.

On the tools side, GitHub Copilot's HydraFusion model picker will extend to local models on Windows, with an experimental preview for the Copilot app, Copilot CLI and Visual Studio Code starting in late October. Microsoft says its MAI Code 1.1 Flash coding model was quantized to 3bit, cutting model size by roughly 80% while keeping coding quality and supporting a 256K context window locally. Windows ML is adding llama.cpp support alongside GPU, NPU and CPU options.

Security is the other half of the story. Because agents act on files, run code and span systems without a person watching, Microsoft is building access limits, identity and organizational management into Windows, and says large-scale management may require additional services. How smoothly those controls work with existing IT practices is likely to shape how quickly enterprises let agents loose on real work.