NVIDIA releases Nemotron-Labs-TwoTower, an open-weight diffusion language model on a frozen autoregressive backbone

Nearly every language model today writes text one token at a time, like a typewriter. Diffusion language models take the image-generator approach instead: they start from a noisy draft and refine it in parallel, producing many tokens per step. The potential upside is speed; the historical downside has been text quality.

NVIDIA is trying to get the best of both worlds. Rather than training a diffusion model from scratch, TwoTower reuses an already-trained autoregressive model (Nemotron-3-Nano-30B-A3B) and freezes it: one tower keeps reading and understanding context exactly as before, while a second, purpose-trained twin tower generates text in parallel using diffusion.

The practical outcome, according to the published numbers: the model keeps almost all of the original quality (98.7%) while producing text at 2.42 times the throughput on the same hardware. For anyone running inference services, the same GPU serves more than twice the requests.

The weights are freely downloadable from Hugging Face under the NVIDIA Nemotron Open Model License, so anyone can verify, benchmark, and integrate them. It is a research-grade release rather than a finished product, but it points in a clear direction: making parallel generation a realistic option for text models too.