They actually have a performance edge, but they aren't well suited to chat models because you can't do caching of past states like with decoder-only models