They are trained from scratch on potentially different datasets. Their architectures are very similar though.