“The primary capability lost when using smaller language models compared to larger ones is reliability.”