Reducing numerical precision (e.g., FP16→INT8/4) to shrink models and speed inference.