Microsoft is loosening a "hard line" it previously drew around Copilot+ PCs, allowing more Windows 11 devices to run native AI workloads if they have the right GPU.The latest update shows that systems with an NVIDIA GeForce RTX 30 series or newer graphics card with at least 6GB of video memory will support Windows' local language model API. On the surface, this is just a small change for developers, but it hints that Microsoft is rethinking whether to tie its native AI capabilities tightly under the Copilot+ brand.

When Copilot+ PC was officially released on June 18, 2024, Microsoft’s message was very clear: dedicated AI hardware is a requirement. Some of the defining features of this type of device are built-in neural processing units (NPUs), as well as basic configurations such as 16GB of memory and solid-state drives. Among them, NPU is deliberately shaped as the key to unlocking Windows’ native AI capabilities.

However, NPUs aren’t the only ones capable of handling AI workloads. Modern GPUs in particular are built for massively parallel computing and have long been used to run machine learning models. In practice, for many AI workloads, GPUs can often provide higher throughput capabilities than current NPUs, often at the cost of higher power consumption.

Prior to this adjustment, Microsoft had been limiting most built-in AI capabilities to devices with NPUs. This prevents many PCs with sufficient computing power and relying solely on GPUs from using local text and image generation, as well as a series of AI tools such as Windows Recall. Today, this gap is beginning to be bridged. Microsoft confirmed in an updated technical document and GitHub post that developers can now run the language model API on non-Copilot+ PCs with supported GPUs.

In its introduction, Microsoft calls this capability "Language Model APIs Run on GPUs (Experimental)," noting that these APIs can now run on non-Copilot+ PCs with supported GPUs, bringing native language model capabilities to a wider range of Windows 11 devices. The official also clarified that currently supported hardware includes NVIDIA GeForce RTX 30 series and newer products equipped with more than 6GB of video memory.

At the current stage, this capability is still mainly at the developer level and is not directly open to ordinary end users. To call these APIs, you need to develop or use an application that integrates the Windows AI Framework. However, this has laid the foundation for the massive expansion of native AI capabilities to more Windows devices.

At the heart of this framework is a small local language model called Phi Silica. Rather than being pre-installed on all systems, Phi Silica is distributed on demand via Windows Update: the model is downloaded only when an app requests it. Once the model is installed, it can be run on local hardware. When an available GPU is detected, the GPU will be used first to accelerate inference.

The currently disclosed functions mainly focus on text-related tasks. Through the Windows.AI.Text API, apps can perform operations such as content summarization, text rewriting, converting text into a structured format, and generating prompts. From a user perspective, these capabilities are similar to the experience provided by cloud AI tools, except that the calculations are completed entirely locally.

Running locally brings some practical advantages. By reducing reliance on cloud computing power, the system response speed is expected to be improved. At the same time, data does not need to be uploaded to external servers, which helps the data stay within the local machine. This model is potentially attractive for both developers and enterprise users in terms of latency, bandwidth costs, and privacy compliance, which could impact how they adopt AI capabilities.

It should be pointed out that this opening does not mean that the Copilot+ system is fully “unlocked”. Some of the more visible Copilot+ features, such as Windows Recall and Click to Do, are currently still tied to NPU-powered systems. As it stands, GPU support is primarily limited to the language model API layer, rather than full integration across the entire AI experience.

Although limitations remain, the trend is already clear: Microsoft no longer sees the NPU as the only entry point for Windows native AI. Allowing the GPU to handle this portion of the workload significantly expands the range of compatible hardware and weakens the "native AI-only" image that Copilot+ PC had at launch.