Does it make sense for a single model to be used for all on device LLM tasks or for each app to provide its own customized one?

My gut feel is the former but not sure if that's actually true.