A specific quantization method that divides model weights into groups and converts each group to lower-precision integers, balancing efficiency with accuracy.