A quantization approach that reduces precision only for model weights while keeping activations at full precision, balancing memory savings with minimal accuracy loss.