The token-by-token responses are probably API, while the "instant" loads seem to be cached.

Probably using groq based on speed of response