Nvidia’s Llama-3.1-Minitron 4B is a small language model that punches above its weight