Distillation
Training a smaller model on the outputs of a larger one, to get most of the quality at a fraction of the cost and latency.
The large model plays the teacher: it answers, and the small model learns to answer the same way. On a narrow task the gap can close almost completely, which is why a distilled model often beats a general one of the same size.
It is the standard move when a feature works but costs too much. Narrow the task first, collect real examples, then measure the small model against the large one on your own evals rather than on a public score.