They have added a small local-only model that runs directly on your machine to try and do smarter autocompletion. The model only runs if your machine is considered powerful enough to provide a sensibly fast experience. Nothing ever leaves your computer. I was equally confused, then I found all these details somewhere in their documentation.
