Thats the innovation of the “chain of thought” models like OpenAI’s o1 and now this deepseek model, it narrates an internal dialogue first in order to try and create more consistent answers. It isnt perfect but it helps it do things like logical reasoning at the cost of taking a lot longer to get to the answer.