NVIDIA Releases Nemotron-Labs-TwoTower, an Open-Weight Diffusion Language Model

The model pairs a frozen Nemotron-3-Nano-30B-A3B autoregressive context tower with a trainable diffusion denoiser. Trained on about 2.1T tokens, it retained 98.7% of the autoregressive baseline's quality with 2.42 times the wall-clock generation throughput. It supports masked diffusion, simulated autoregressive, and standard autoregressive inference, and has about 60B combined parameters with 3B active per tower. The weights are on Hugging Face under the NVIDIA Nemotron Open Model License, permitting commercial use.

Visit original source