IIRC quantizing small models causes a higher relative drop in the metrics.