A quantization strategy that uses 4-bit precision for some weights and 8-bit precision for others, balancing memory savings with accuracy.