More data, more fine-tuning. Over 2B more tokens of training.
Umm, devs? Could you add something here?