A quantization approach that applies different compression settings to different channels or groups within a model layer, helping preserve quality better than applying the same compression uniformly.