Presumably one of the benefits of running your own model is that you can feed extra data into it via training rather than purely through inference? I.e. if you're a software company you could fine-tune it on your codebase, improving its answers without increasing inference time?