The details on how exactly they may have used it to train their model is vague. I believe transfer learning or knowledge distillation are valid techniques based on the inference from other models.