“An optimization technique where model weights are bundled into the container to prevent separate downloads (e.g., from Hugging Face), reducing cold start times.”