>But you _will_ have to accept lower quality outputs compared to todays' frontier models.

I'm curious how much lower quality we're talking about here. Most of the work I ever get an LLM to do is glue-code, or trivial features. I'd expect some fine-tuned Codestral type model with well focused tasks could achieve good performance locally. I don't really need worlds-leading-expert quality models to code up a hamburger menu in a React app & set the background-color to #A1D1C1.