The Windows Developer Blog highlights the launch of Distilled DeepSeek R1 models for Copilot+ PCs, enhancing AI capabilities directly on devices. With optimized versions for Qualcomm and Intel processors, developers can create efficient AI applications using the AI Toolkit. These models leverage advanced NPU technology for improved performance and reduced power consumption, enabling seamless local interactions.

Unleashing the Power of AI on Copilot+ PCs

AI technology is rapidly evolving, and the latest development from Microsoft is turning heads. The introduction of Distilled DeepSeek R1 models on Copilot+ PCs marks a significant leap in on-device AI capabilities.

What’s New?

The integration of DeepSeek R1 models into the Windows Copilot Runtime is groundbreaking. This new feature allows developers to run NPU-optimized versions of DeepSeek directly on their PCs. Initially, support will roll out for Qualcomm Snapdragon X, followed by Intel Core Ultra 200V. >

“With the availability of cloud-hosted DeepSeek R1 on Azure AI Foundry, we’re bringing powerful AI capabilities right to your desktop.”

The first model, DeepSeek-R1-Distill-Qwen-5B, is available now. Soon, variants with 7B and 14B parameters will follow. These models are designed to run efficiently on-device, taking full advantage of the powerful Neural Processing Units (NPUs) in Copilot+ PCs.

Major Updates in AI Toolkit

Developers can easily access these models via the AI Toolkit in Visual Studio Code. Simply download the extension, and you can start experimenting with DeepSeek. The models will be available in the ONNX QDQ format, allowing seamless integration into your projects. >

“Experimenting with the model is as simple as opening the Playground, loading the model, and sending it prompts.”

This streamlined process empowers developers to test and deploy AI applications effectively. The Playground feature allows for quick iterations, ensuring that your AI models are ready for real-world applications.

What’s Important to Know?

The NPU’s efficiency is a game-changer. It allows for semi-continuous AI services, enhancing user experiences. The DeepSeek models leverage advanced techniques like low bit-rate quantization to optimize performance without draining battery life. The Qwen 5B model uses a sliding window design for rapid response times. This innovation delivers a time to first token of just 130 ms, making interactions feel instantaneous. Moreover, the combination of low-bit quantization and hardware optimizations ensures that the models maintain their reasoning capabilities while being power-efficient. In conclusion, the rollout of Distilled DeepSeek R1 models on Copilot+ PCs signifies a new era of AI development. With these powerful tools at your fingertips, the possibilities for AI applications are endless. Get ready to dive into the future of AI with Microsoft’s latest innovations!

DeepSeek R1 models are designed for local deployment on Copilot+ PCs, enhancing on-device AI capabilities.

The models utilize advanced NPU technology for efficient inferencing, enabling continuous AI service execution.

Developers can easily experiment with models using the AI Toolkit VS Code extension and Azure AI Foundry.

Optimizations include low-bit quantization and a sliding window design for improved speed and memory efficiency.

The NPU-optimized models maintain reasoning abilities while delivering faster response times and lower power usage.

From the Windows Blog

---