NVIDIA Releases Nemotron-Labs-TwoTower: an Open-Weight Diffusion Language Model Built on a Frozen Autoregressive Nemotron-3-Nano-30B-A3B Backbone

NVIDIA has released an open-weight diffusion language model that generates text faster than traditional autoregressive models by producing multiple tokens in parallel and refining them iteratively rather than generating one token at a time. The model uses a two-tower architecture where a frozen context tower maintains the original language model's knowledge while a trained denoiser tower refines noisy text blocks, achieving 2.42 times higher generation throughput while retaining 98.7% of the baseline model's benchmark quality. This approach addresses a fundamental bottleneck in text generation where autoregressive decoding commits exactly one token per step, whereas the new model can commit multiple tokens per step during the refinement process. The technology offers practical applications including faster batch text generation for synthetic data production and the ability to tune the trade-off between generation speed and output quality from a single checkpoint.