Training a high-quality Japanese model from scratch requires hundreds of billions of Japanese tokens, whereas incremental training only requires supplementing Japanese corpora on top of existing weights.