why not an inbetween scenario like using a managed inference provider to host your own models?