Microsoft just announced that it will release NPU-optimized versions of DeepSeek-R1, allowing it to take advantage of AI-optimized hardware found in Copilot+ PCs. According to the Windows Blog, the feature will first arrive on Qualcomm Snapdragon X PCs, to be followed by Intel Core Ultra 200V (Lunar Lake) and other chips. The initial release will feature DeepSeek-R1-Distill-Qwen-1.5B, which an AI research team from UC Berkeley has discovered is the smallest model that delivers correct answers, but larger models featuring 7 billion and 14 billion parameters will arrive shortly thereafter.
DeepSeek’s optimizations meant that it needed 11x less compute versus its Western competitors, making it a great model to run on consumer devices. However, it also uses Windows Copilot Runtime so developers can use on-device DeepSeek APIs within their apps.
Furthermore, Microsoft claims that this NPU-optimized version of DeepSeek will deliver “very competitive time to first token and throughput rates, while minimally impacting battery life and consumption of PC resources.” This means that Copilot+ PC users can expect the power and performance of competing models like Meta’s Llama 3 and OpenAI’s o1 while ensuring that the devices it’s installed on still offer great battery life.
That said, DeepSeek’s availability on Copilot+ PCs is geared more toward programmers and developers instead of consumers. Perhaps Microsoft is using it to encourage them to build more apps that would take advantage of AI PCs as many people still don’t see the need for it and market research suggests users only purchase these devices because they’re the only available option nowadays.
Another thing that got us curious is Microsoft’s preferential treatment for Qualcomm Snapdragon X PCs at this time. While it launched the Copilot+ branding with these chips last July, the latest mainstream Intel and AMD laptops now also have built-in NPUs. AMD has even released instructions on how users can run it on Ryzen AI CPUs and Radeon GPUs, with the company even claiming that the RTX 7900 XTX runs DeepSeek better than the RTX 4090.
Whatever the case, we’re still excited about the possibilities that DeepSeek unlocks for AI. Since it’s open source, nearly anyone can download it and run it locally, allowing others to build upon the advancements and optimizations the original model has put into place.
