It really isn't, you can improve by distilling a weaker model