seems like they use the normal transformer architecture versus deep fold's more specialised machine-learning approaches.