I think some of these 13B models may only be pretrained and may not have benefited from RLHF. So it’s not clear how much of the difference is due to parameter count.