kimi seems routing the input into different models, hard problems tend to take much longer time and chain-of-thought than easy problems even if the selected model is k3.1 pro