But are these llms worth their salt?

They're not unless you curve the grading because they're running locally.

Which some people do, but I don't think the average person asking this question does (and I don't)

With 128GB of memory they can have real world use cases. But they won’t be as good as SoTA hosted models.