There are a couple of “thinking” models that output text just like that. They chew the query over and over extending the prompt to improve accuracy