Detecting humor is a necessary step towards language understanding. However, work on computational humor for Portuguese is still limited. For this language, we tackle the task of pun location. With a corpus of annotated punning texts, we fine-tune available encoder models for labeling words in context as punning or not. We achieve an F1 of 0.75 with a BERT-based model and further improve precision with post-processing. Moreover, we show that a model trained for pun location can be used for pun detection as well, performing close to a model specifically trained on the latter task, but with the advantage of identifying the pun words, thus contributing to explainability.