Quantization compresses model weights from high-precision floating point numbers such as FP32 to low-precision representations such as INT8 or INT4, reducing memory footprint and computational requirements