NVIDIA released Nemotron-TwoTower-30B-A3B-Base-BF16, an unusual diffusion-based language model built on Nemotron 3 architecture. Represents alternative approach to traditional autoregressive language modeling.