Comment by largbae

Yes, to make training data for other models. You can distill to get better at something the source model is good at, perhaps it has more training data in a language that you want to support better. Also large open weight models are a great source of training data for smaller models.