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.