Despite being primarily trained on text, these models achieved remarkable compression rates on image and audio data.