Pre-baking model weights into a Docker image can eliminate the runtime model download step, but the loading time from disk to VRAM still exists and is typically billed.