NVIDIA releases Nemotron-Labs-TwoTower, an open-weight text model built on diffusion

NVIDIA has published Nemotron-Labs-TwoTower, a text model with open weights. It is a diffusion language model built on Nemotron-3-Nano-30B-A3B and designed to speed up text generation.

What limits autoregressive models

Autoregressive (AR) architectures underpin GPT, Llama, Gemini and most other LLMs. The principle is simple: each next token is predicted from all the previous ones, and the process is strictly sequential. This is elegant at training time, but it creates a concrete problem for production inference.

The next token cannot be computed until the previous one is finished. Adding GPU accelerators to a cluster does not remove this limit: it is built into the architecture at the level of the computational graph, and decoding is sequential by definition. For long responses, users wait proportionally longer, and the cost of a single token hits a hard floor as you scale. For providers handling billions of requests a day, that means direct, ongoing operating costs. This is why faster inference is one of the main research directions in the industry, alongside smaller models and quantization.

Discrete diffusion language models offer an alternative mechanism. Instead of decoding step by step, they iteratively refine the entire output block over several steps. This opens the potential for generating several tokens in parallel in a single pass, and with it a fundamentally different throughput profile.

How the TwoTower architecture works

The two-tower design combines the AR and diffusion approaches in a single model:

• AR backbone: pretrained Nemotron-3-Nano-30B-A3B (30 billion parameters, frozen)
• Diffusion head: a trainable component on top of the frozen backbone
• Open weights: NVIDIA Nemotron Open Model License
• Release date: July 1, 2026

Freezing the AR backbone is a key architectural decision. Rather than training a diffusion model from scratch, NVIDIA uses the pretrained AR base as a fixed source of contextual language representations. Only the diffusion component is trained, which lowers the compute cost per experiment and reduces the data needed for adaptation. Choosing Nemotron-3-Nano-30B-A3B as the backbone also makes the work easier to reproduce: other teams can repeat the experiment starting from the same publicly available checkpoint.

Why diffusion for text is a hard problem

Diffusion models became the standard for image generation: Stable Diffusion, Midjourney and DALL-E 3 all work on this principle. Transferring it to text is far harder. Pixels live in a continuous numerical space, whereas tokens are discrete. Standard Gaussian noise cannot be applied to discrete objects, so text requires dedicated discrete diffusion processes.

This field is developing actively but is still young by industry standards. Earlier work, including MDLM and SEDD, showed competitive results on language benchmarks, but a quality gap with the best AR systems remained. NVIDIA's two-tower approach is an attempt to remove that trade-off: take the strong language representations of an already trained AR model and add a diffusion mechanism to them, without losing the accumulated knowledge of syntax, semantics and contextual dependencies.

Open weights are especially valuable for the academic community: researchers can reproduce the architecture, measure the actual throughput gain on their own tasks and propose improvements on top of the published checkpoint.

What it means

Nemotron-Labs-TwoTower is a practical step toward faster LLM inference without replacing hardware. As a leading GPU supplier to the AI market, NVIDIA has an interest in widening the applicability of language models, including by lowering inference costs. If the hybrid AR-plus-diffusion approach proves itself in real load scenarios, both in generation quality and in actual throughput gains, it may influence architectural choices when developing the next generation of language systems.

What is Nemotron-Labs-TwoTower?

It is a text model from NVIDIA that uses a diffusion architecture instead of the traditional autoregressive approach. The goal of the development was to eliminate the bottleneck in text generation speed.

Can the NVIDIA Nemotron model be used in Russia?

Yes. The model is distributed under the open NVIDIA Nemotron Open Model License, which allows its use in any country.