Researchers found models with about 3 billion parameters perform almost as well as those with up to 14 billion, potentially reducing the computing power needed for brain research.