A specific quantization scheme where weights are stored in 4-bit precision while activations remain in 16-bit precision, balancing memory savings with accuracy.