Recognizing Textual Entailment, also known as Natural Language Inference recognition, aims to identify if it is possible to infer the meaning of a text from another fragment of text. In this work, we investigate the use of multilingual models, through BERT, for recognizing inference and similarity in the ASSIN 2 dataset, an entailment recognition and sentence similarity corpus for Portuguese. We also investigate possible features that could enhance the results, such as similarity scores or WordNet relations. Our results show that a multilingual pre-trained BERT model may be su.