4×

smaller weights

32-bit floats → 8-bit integers

75%

less memory

per parameter

0

network calls

inference stays on device