NVIDIA has released Nemotron-TwoTower-30B-A3B-Base-BF16, an unusual diffusion-based language model built from the Nemotron 3 Nano 30B-A3B backbone. Instead of generating strictly one token at a time, it uses a frozen autoregressive context tower plus a diffusion denoiser tower that iteratively fills blocks of tokens in parallel. NVIDIA says its default mask-diffusion setup retains 98.7% of the autoregressive baseline’s aggregate benchmark quality while reaching 2.42× its wall-clock generation ...