NVIDIA's Nemotron TwoTower Is a Diffusion LLM That Writes Text in Parallel — 2.42× Faster

Almost every large language model you have used writes the same way: one token at a time, left to right, each word waiting on the one before it. That serial process is also the ceiling on how fast a model can generate. NVIDIA just released an open-weight model that breaks the mold — Nemotron-Labs-TwoTower, a diffusion language model that generates text in parallel and refines it in a few iterative passes, the way image diffusion models paint a whole picture at once rather than pixel by pixel.

The payoff is speed. TwoTower reports 2.42× higher wall-clock generation throughput than a comparable autoregressive model while retaining 98.7% of its aggregate benchmark quality. In other words, it gives up almost nothing in accuracy to run more than twice as fast — the kind of trade that matters enormously when inference cost, not training, is what dominates the bill for anyone serving a model at scale.

What makes the release clever is how it was built. Rather than train a diffusion model from scratch, NVIDIA instantiated TwoTower on Nemotron-3-Nano-30B, an existing open-weight hybrid backbone that interleaves Mamba-2, self-attention, and mixture-of-experts layers. The "two towers" both start as copies of that same checkpoint — but only one, the denoiser, is trained, while the other stays frozen as a context model. The denoiser saw just ~2.1 trillion tokens, a small fraction of the backbone's 25-trillion-token pretraining. Diffusion, in effect, was bolted onto a strong autoregressive model instead of replacing it.

That approach hints at a broader shift. For years, "diffusion" meant images and video while "autoregression" meant text, and the two rarely mixed. TwoTower is part of a wave of research — including NVIDIA's own tri-mode Nemotron work — arguing that the boundary is artificial, and that discrete diffusion can deliver the parallelism text generation has always lacked. If block-wise diffusion decoding holds up on longer, harder tasks, it reframes a core assumption about how LLMs should be built.

It also lands as an open-weight release under NVIDIA's Nemotron Open Model License, which matters as much as the benchmarks. By publishing the weights, NVIDIA hands researchers and startups a working diffusion LLM to poke at, fine-tune, and deploy — accelerating a line of work that could make the next generation of models cheaper to run. In a year defined by compute shortages and inference bills, a technique that squeezes 2.42× more throughput out of the same hardware is not a curiosity. It is exactly the kind of efficiency the whole industry is hunting for.