More and more we are hearing about small language models with hundreds of millions or sub 10 billion parameters that are highly accurate and consume substantially less energy and cost less per token.