Details
- NVIDIA AI introduces Nemotron-Labs-TwoTower, a diffusion language model built on the Nemotron-3-Nano-30B-A3B backbone, effectively splitting a 30B model into two coordinated towers.
- The architecture uses a frozen autoregressive context tower to hold and process the prompt and committed tokens, while a trainable diffusion denoiser tower generates and refines blocks of tokens in parallel.
- TwoTower operates as a block-wise autoregressive diffusion model, iteratively denoising masked token blocks instead of emitting one token at a time, improving generation throughput on modern GPUs.
- The released checkpoint includes both towers and is available as open weights under the NVIDIA Nemotron Open Model License, making it suitable for commercial deployment.
- NVIDIA reports that TwoTower retains about 98.7% of the baseline autoregressive model’s aggregate benchmark quality while delivering roughly 2.42× higher wall-clock generation throughput at its default operating point.
- The model is hosted on Hugging Face, with technical details provided in an accompanying research paper describing the TwoTower design and its adaptation from Nemotron-3-Nano-30B-A3B.
- This release extends the Nemotron-Labs Diffusion line of models, which explore diffusion-based parallel decoding and hybrid AR–diffusion approaches for faster, more efficient text generation.
Impact
Nemotron-Labs-TwoTower signals NVIDIA’s push to make diffusion-based language models practical for high-throughput inference, narrowing the efficiency gap with leading autoregressive systems while keeping quality nearly intact. By shipping open, commercially usable weights and integrating cleanly with the Nemotron backbone, NVIDIA strengthens its position in the enterprise and research LLM stack and pressures rivals to offer similarly efficient, parallel-friendly decoding schemes.
