Yeah, for short inputs the models are not that precise, but for long and complicated ones they work quite good, probably NLLB has been trained with very few short texts.