As I said elsewhere, Deepseek injects Chinese characters into responses. Anecdotally, that seems to happen when the context gets longer. That suggests that they're primarily trained in Chinese and I would expect them to use fewer tokens for Chinese than English.
As I said elsewhere, Deepseek injects Chinese characters into responses. Anecdotally, that seems to happen when the context gets longer. That suggests that they're primarily trained in Chinese and I would expect them to use fewer tokens for Chinese than English.