“Chinese AI labs likely use both distillation from leading closed models and distillation from their own domain-specific models in their post-training processes.”