They are based on llama2 architecture which performs a lot worse.