The crux is how big the L is in the local LLMs. Depending on what it's used for, you can actually get really good performance on topically trained models when leveraged for their specific purpose.