If their models truly are reaching parity with new western models as many claim, then it can't be from distillation alone, they must be gathering their own datasets too. It takes months to train a new model.

Also distillation only really saves you the data collection and preparation (categorization). Training is still expensive, as is inference. It's likely architectural changes that are making their inference cheaper (MoE vs dense models for one), not sure if they've gotten any good methods for making training cheaper.