Microsoft Overhauls Windows 11 AI Strategy, Shifts to On-Device Large Models

Microsoft is making a significant adjustment to its artificial intelligence strategy for Windows 11, planning to gradually move away from reliance on the physical Copilot key on keyboards and the web-based sidebar, and instead shift toward running large AI models directly on users' personal computers. This transition signals that the focus of AI integration with the operating system is moving from surface-level interaction entry points to the underlying computing architecture of the OS.

Over the past two years, Microsoft has pursued an intensive software and hardware bundling strategy to promote Copilot, including working with OEM partners to add the first new physical key to PC keyboards in nearly three decades—the Copilot key—and embedding AI invocation buttons in core areas of Windows 11 such as the taskbar and File Explorer. However, these designs have sparked controversy after reaching the market. Many users have pointed out that, at its current stage, Copilot is largely a cloud-driven web wrapper, which responds slowly on poor networks and fails to integrate deeply with the system's underlying workflows.

In response, Microsoft has outlined an 'edge-first' technical roadmap in its latest system plans and developer programs. The core of this approach is to offload AI model inference and computation directly to local devices. Leveraging the high-compute neural processing units (NPUs) in next-generation PC hardware, as well as system platforms that support large-capacity, high-bandwidth unified memory architectures (such as the Project Zenith ecosystem for professional developers), Windows 11 is being transformed into a high-performance environment capable of running models with billions or even tens of billions of parameters locally, without relying on cloud servers.

Running AI models locally eliminates network latency caused by data round-trips to the cloud and brings privacy and security benefits. System-level sensitive operations, screen content awareness, and local document indexing can all be processed in a closed loop within the device, without uploading personal data to remote cloud servers. For developers, on-device models mean they can move away from cloud-based usage billing and build AI applications with high responsiveness and security features in offline or low-connectivity environments.