The Nemotron 3 models use a hybrid latent mixture-of-experts (MoE) architecture to run AI systems with improved speed and ...