Researchers are testing small language models to see if they can outperform the large models they are based on.