In January, Microsoft introduced its plans to convey NPU-optimized variations of the DeepSeek-R1 mannequin on to Copilot+ PCs powered by Qualcomm Snapdragon X processors. In February, DeepSeek-R1-Distill-Qwen-1.5B was first made obtainable in AIToolkit for VSCode.

At present, Microsoft introduced the supply of DeepSeek R1 7B and 14B distilled fashions for Copilot+ PCs through Azure AI Foundry. The power to run 7B and 14B fashions domestically on Copilot+ PCs will allow builders to construct new sorts of AI-powered purposes that weren’t attainable earlier than.

Because the fashions run on NPUs, customers can count on sustained AI compute energy with much less impression on their PC battery life and thermal efficiency. Additionally, their CPU and GPU can be obtainable to carry out different duties.

Microsoft highlighted that it used Aqua, an inner computerized quantization instrument, to quantize all of the DeepSeek mannequin variants to int4 weights. Sadly, the mannequin token pace is kind of low. Microsoft is reporting solely 8 tok/sec on the 14B mannequin and near 40 tok/sec with the 1.5B mannequin. Microsoft talked about that it’s engaged on additional optimizations to enhance the pace. As Microsoft continues to optimize efficiency, the impression of those fashions on Copilot+ PCs is anticipated to develop considerably.

builders can obtain and run the 1.5B, 7B, and 14B variants of DeepSeek fashions on Copilot+ PCs through the AI Toolkit VS Code extension. The DeepSeek mannequin is optimized within the ONNX QDQ format and is downloaded instantly from Azure AI Foundry. These fashions can even be coming to Copilot+ PCs powered by Intel Core Extremely 200V and AMD Ryzen processors sooner or later.

This transfer from Microsoft signifies a push in direction of extra highly effective on-device AI capabilities, opening up new potentialities for AI-driven purposes.