Microsoft recently low-key updated the document stating that as long as the device is equipped with a discrete graphics card that meets the conditions, even if it is not a Copilot+ PC, it can run the local language model API on Windows 11. This is seen as an important step in weakening the exclusive advantage of Copilot+ PC. According to the latest instructions, eligible hardware includes NVIDIA GeForce RTX 30 series and newer models, and requires 6GB of video memory and above.
This means that some high-performance PCs that do not originally meet the Copilot+ PC standard will be able to obtain system-level local AI capabilities through GPUs, instead of relying entirely on new models with high computing power NPUs.
Copilot+ PC was officially unveiled on June 18, 2024. Microsoft packaged it as the “only way” to run local AI at the time, and used it as an important selling point to promote PC upgrades. According to Microsoft's original definition, for a device to be labeled as a Copilot+ PC, it needs to be equipped with at least 16GB of memory, a solid-state drive, and an NPU (neural processing unit) that can provide no less than 40 TOPS of computing power. NPU is designed to run AI models efficiently, emphasizing energy efficiency, while GPU is good at large-scale parallel computing and has stronger raw computing power. However, under the previous strategy, Microsoft limited local AI functions such as Windows Recall and "Click to Do" to Copilot+ PCs equipped with NPUs, resulting in many GPU computers with equally powerful hardware performance being unable to use these built-in AI capabilities.
This turn comes from an update note from Microsoft on GitHub. Microsoft confirmed in the document that the language model API can already be run on GPU on non-Copilot+ PCs, as long as two conditions are met: first, you have an RTX 30 series or newer graphics card, and second, the graphics card has at least 6GB of video memory. "The language model API now runs on non-Copilot+ PCs with supported GPUs, bringing native language model capabilities to a wider range of Windows 11 devices," Microsoft said in the note. Supported hardware currently includes the NVIDIA GeForce RTX 30 series and later product lines, and requires 6GB of video memory and above. In practical terms, this change extends the scope of local language models from new Copilot+ PCs to the large number of high-performance Windows 11 devices already in use.
From a developer's perspective, the language model API provides a set of Windows.AI.Text interfaces that can be called locally, supported by a small language model called "Phi Silica". The model has previously appeared in products like Microsoft Edge, for features such as "Rewrite using AI." According to Microsoft's settings, ordinary Windows PCs currently do not pre-install local AI models by default. Only users who purchase Copilot+ PCs will have relevant content preset in the system. Under the new scheme, if a user runs an application that supports the local language model API, the system can download the Phi Silica model through Windows Update and run it locally on a qualified GPU.
Based on the currently announced interface capabilities, the local AI functions that developers can call in their applications include: intelligent text formatting using the Windows.AI.Text interface, text summary (TextSummarizer), rewriting (TextRewriter), text to table (TextToTableConverter), and general prompt word generation, etc. In other words, as long as the hardware meets the standards, users will be able to obtain a partial text processing experience similar to ChatGPT natively in Windows applications, and the calculation process is completely completed locally, which has more advantages in privacy protection than Copilot or ChatGPT that rely on cloud models.
It should be noted that Microsoft currently only liberalizes the scope of GPU application at the language model API level. Judging from the available information, some AI functions in applications such as Windows Recall, "Click to Do" and MS Paint are still only available to Copilot+ PCs with built-in NPUs. In other words, this adjustment is more about "unlocking" a wider range of hardware platforms for local AI capabilities such as text and prompts, rather than completely removing the threshold for NPUs in the Copilot+ ecosystem. Nonetheless, as local AI capabilities gradually extend to traditional high-performance GPU PCs, the selling point of Copilot+ PC as "the only one that can run AI locally" has been significantly weakened, which has also led to outside speculation about whether Microsoft will further relax or even cancel the mandatory requirements for NPUs in the future.