“Imprecise” language models are smaller, speedier—and nearly as accurate