Text model

Phi Silica

Phi Silica is Microsoft’s on-device small language model (~3.3B params) built to run locally on Copilot+ PC NPUs. It’s pre-tuned, 4-bit–quantized, streams fast (≈230 ms first token, up to ~20 tok/s), and currently offers a 2K context (with 4K coming). Available to developers via the Windows App SDK’s Phi Silica APIs.

Type
Text
Weights
Open