Microsoft Unveils Phi-Silica: A 3.3B Parameter Model Designed for Copilot+ and Advanced PC NPUs

Microsoft is ramping up its investment in small language models (SLMs). During its Build developer conference, the company announced the general availability of its Phi-3 models and previewed Phi-3-vision. Following the news about Microsoft’s Copilot+ PCs, the company is introducing Phi-3-Silica, an SLM specifically designed for the powerful Neural Processing Units (NPUs) in these devices.

Phi-3-Silica will be included in all Copilot+ PCs available starting in June. This model is the smallest in the Phi lineup, featuring 3.3 billion parameters.

According to Microsoft, the first token latency for Phi-3-Silica is 650 tokens per second and consumes only 1.5 Watts of power, ensuring it remains lightweight and efficient. This allows the PC's CPU and GPU to focus on other tasks. Additionally, its token generation can leverage the NPU’s KV cache, enabling the CPU to produce around 27 tokens per second.

A Microsoft spokesperson emphasized that Phi-Silica is notable as the first locally deployed language model for Windows. Optimized for the Copilot+ PCs' NPU, it brings fast local inference to users' devices. This development signifies an important step in equipping third-party developers with advanced AI tools optimized for Windows, enhancing productivity and accessibility within the ecosystem. Users can expect innovative experiences from both first-party and third-party developers this fall.

Phi-Silica is the fifth addition to Microsoft's Phi-3 series, which includes Phi-3-mini with 3.8 billion parameters, Phi-3-small with 7 billion parameters, Phi-3-medium with 14 billion parameters, and Phi-3-vision with 4.2 billion parameters.