A model version specifically tuned to run efficiently on particular hardware (like NVIDIA GPUs), taking advantage of that hardware's strengths.