NVIDIA has released Nemotron 3 Nano, a hybrid Mamba-MoE model designed to cut inference costs by 60% and accelerate agentic ...