At a Glance
| Vendor | NVIDIA |
| Family | Nemotron series |
| Launched | NVIDIA released Nemotron-Labs-TwoTower on July 1, 2026 as an open-weight diffusion language model under the NVIDIA Nemotron Open Model License. The accompanying paper is arXiv 2606.26493. A tri-mode weights update followed on approximately July 10. |
| Context window | Inherited from the Nemotron-3-Nano-30B-A3B backbone, which interleaves Mamba-2, self-attention, and mixture-of-experts layers. |
| Pricing | Free to self-host under the NVIDIA Nemotron Open Model License. This is a research-grade release aimed at inference-efficiency work rather than a hosted commercial product. |
| Access channels | Hugging Face open weights, the arXiv technical report, and NVIDIA's Nemotron model catalogue. |
Notable Benchmarks
TwoTower retains 98.7 percent of autoregressive quality at 2.42x throughput, measured at gamma 0.8, S 16, on 2x H100. The released checkpoint ships both towers at roughly 60B total parameters, with approximately 3B active per token per tower. The denoiser was trained on about 2.1 trillion tokens, a fraction of the backbone's 25 trillion.
Strengths
The architecture is the point. TwoTower splits diffusion into a frozen autoregressive context tower and a trained denoiser tower, which means any lab holding an existing autoregressive checkpoint can bolt on diffusion-based parallel generation by training only a second network on a fraction of the original data budget. That is a far cheaper path to parallel decoding than training a diffusion LLM from scratch.
Limitations
It is a research release, not a production frontier model, and absolute quality tracks the 30B-class backbone rather than the frontier. The 2.42x figure is configuration-specific and does not transfer unchanged to other hardware or settings.
Brand-Visibility Implications
Diffusion language models generate tokens in parallel rather than strictly left to right, which changes how an answer is assembled and therefore potentially how brand mentions are placed within it. Position within an AI answer is a measured driver of click-through and recall, and the GEO literature's position-adjusted word count metric assumes sequential generation. If parallel decoding reaches production scale, that assumption needs re-testing. This is the earliest open-weight release worth watching on that question. See GEO academic papers and Nemotron 3 Nano Omni brief.
How Presenc AI Tracks This Model
Presenc AI monitors brand visibility on NVIDIA's Nemotron series as part of continuous multi-platform AI visibility tracking. We sample NVIDIA Nemotron-Labs-TwoTower across representative prompt sets daily, compare against competitor performance on the same prompts, and flag material mention-rate changes so brand teams can respond quickly when AI representation shifts.