It should be able to solve tags with Deepseek R1 local llama.cpp, so that thinking can collapsed or hidden.