This is a big part of it. CPU-only models are much easier to deploy, and not all developers have GPUs, so the first iteration of the local model will be CPU-only. We might release a GPU version later.