A lot of open weight models don't understand intent well, they'll overfixate on a word in the prompt or just go off the rails trying to do much work.

GLM-5.2 actually has really good intent understanding though, on par with GPT-5.5 and Opus from my experience.