Definitely. I tried gemma2:27B model with phrases like "translate the following sentence to language X" and it even failed to understand the task and spat out completely irrelevant things, like math formulas.

OTOH, smaller model did it perfectly.