Microsoft Brings DeepSeek R1 AI Models to Copilot+ PCs

Microsoft has expanded its AI offerings by making DeepSeek R1 7B and 14B distilled models available for Snapdragon-powered Copilot+ PCs via Azure AI Foundry. This move enhances on-device AI performance and reinforces Microsoft's vision for hybrid AI—combining cloud and edge computing.

Key Highlights

✅ Availability

• As of March 3, 2025, the 7B and 14B models are available on Snapdragon-powered Copilot+ PCs.
• Future expansion is planned for Intel Core Ultra 200V and AMD Ryzen-powered devices.

✅ Developer Access & Performance

• Models are accessible via the AI Toolkit VS Code extension and optimized in ONNX QDQ format.
• Performance:
  • 14B model: ~8 tokens/sec
  • 1.5B model (released earlier): ~40 tokens/sec
• Further speed and efficiency optimizations are underway, using QuaRot quantization and sliding window techniques.

✅ Optimized for AI Efficiency

• Uses a hybrid approach balancing CPU, GPU, and NPU workloads:
  • 4-bit block-wise quantization for memory-intensive operations.
  • int4 per-channel quantization with int16 activations for compute-heavy transformer blocks.
• Designed for on-device AI applications that run efficiently without straining system resources.

Why It Matters

This release is part of Microsoft’s broader AI-first Windows strategy, bringing powerful AI capabilities directly to consumer devices.

🔹 DeepSeek R1 models—originally distilled from a 671B-parameter AI—offer on-device reasoning without cloud reliance.
🔹 Microsoft-Qualcomm collaboration highlights Snapdragon’s role in edge AI acceleration.
🔹 Despite OpenAI’s disputes over DeepSeek’s data origins, Microsoft continues to push integration into its ecosystem.

What’s Next?

Microsoft plans to improve performance and expand model support to Intel and AMD Copilot+ PCs, broadening access to AI-driven applications on Windows.

For now, developers with Snapdragon-powered Copilot+ PCs can tap into the 7B and 14B models, marking a major step in bringing AI-driven reasoning to personal devices.