Seems kind of like phi but for writing, the smaller ones are trained with 50B tokens and the largest is only trained with 18B.