Source:

Phi Silica, small but mighty on-device SLM

Phi Silica is designed to run locally on Copilot+ PCs using Neural Processing Units (NPUs).

It delivers low latency, high efficiency, and no cloud dependency, making AI features faster and more private.

🔹 Powerful Yet Efficient

• Runs models with 3–7 billion parameters.
• NPUs offer up to 20x more performance and 100x more efficiency than traditional GPU-based AI workloads.

🔹 Integrated Windows Features

• Powers new Windows 11 features like:
  • Click to Do (Preview)
  • On-device rewrite and summarize in Word and Outlook
  • Recall (Preview) for retrieving past user activity

🔹 Developer Access & Multilingual Support

• Developers get access to the Phi Silica API starting January 2025.
• Supports 4k context length and multiple languages including English, Chinese, French, German, Japanese, and more.

🔹 Privacy and Cost Benefits

• Eliminates the need for cloud-based subscriptions.
• Enhances privacy by keeping AI processing on-device.

🔹 Built for Windows

• Tailored specifically for Windows with turnkey integration—no extra tuning needed.
• Based on a Cyber-EO compliant version of Phi-3.5-mini.