If that aspect is critical to your workflow, it seems like you could run the model of your choice off of hugging face, on GPU hardware under your control, so the model won't shift out from under you.