> But the LLM can think between writing each token

Writing a token is the thinking itself. Thinking models just write some tokens behind the scene, that's the whole difference.