In quantization, the number of weights that share a single scaling factor; smaller groups preserve more precision but use more memory, while larger groups save more memory but may lose detail.