NVIDIA froze half a model to make the other half 2.4× faster. The math is weird. The weights are public.

Language models generate text the same way a person types: one token at a time, left to right, each word waiting on the word before it. That sequential dependency is fundamental to the architecture. It's also an inference bottleneck that every optimization trick in the stack — speculative decoding, KV cache compression, continuous batching — is working around rather than solving.

Diffusion language models are the attempt at solving it directly. Generate all tokens in rough parallel. Iteratively refine. Arrive at the same output in fewer sequential passes.

The problem: every diffusion LM built from scratch has fallen short. Either slower in practice than the theoretical parallelism suggests, or lower quality than a comparable AR model, or both. The architecture is theoretically right. The empirical results haven't followed.

NVIDIA's Nemotron-Labs-TwoTower, released July 2 on HuggingFace, is the most interesting attempt at closing that gap I've seen in the last year. It claims 2.42× higher wallclock generation throughput versus its autoregressive baseline while retaining 98.7% of aggregate benchmark quality. And the trick isn't a novel architecture. It's a frozen old one.

What TwoTower actually does

Most diffusion LMs start from scratch. TwoTower doesn't.

The model takes Nemotron-3-Nano-30B-A3B — a pretrained 30-billion-parameter mixture-of-experts backbone with 3 billion active parameters, trained on roughly 25 trillion tokens — and copies the weights into two separate towers.

The AR context tower is then frozen entirely. No further training. The diffusion denoiser tower is trained from that same starting checkpoint on an additional 2.1 trillion tokens.

The reasoning: the frozen AR tower handles contextual understanding — what the sequence means, what has happened so far, what continuations are coherent. That understanding is expensive to develop (25T tokens of pretraining) and already exists. The diffusion tower inherits it and specializes only in the refinement problem: given a noisy draft of future tokens, denoise toward the correct answer.

Together: approximately 60 billion total parameters. Three billion active per tower.

The denoiser doesn't need to build world knowledge or language understanding from scratch. It just learns to refine. That's why 2.1T tokens of additional training is enough — it's a much narrower task than pretraining.

Source spread

• arXiv 2606.26493 — Nemotron-TwoTower paper — academic. Primary source. The architecture, training procedure, throughput and benchmark methodology. Read the methodology section before citing the quality numbers.
• HuggingFace — nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16 — builder. Weights and model card. The license terms live here.
• MarkTechPost — release overview — builder. Parameter counts, training token counts, throughput figure.
• TechTimes — throughput analysis — skeptic. Notes the "without retraining" framing and what the hardware conditions of the throughput benchmark actually are.
• NVIDIA Research — Nemotron-Labs-Diffusion May 2026 — academic. The preceding tri-mode paper that established the diffusion-AR hybrid research direction.

Pros & cons

What's real:

• The architecture insight is sound. Freezing the AR backbone means the denoiser inherits rich pretrained representations without having to develop them independently. The 2.1T training token count for the denoiser (vs 25T for the backbone) is evidence this works — you're not doing the hard part twice.
• 2.42× throughput is significant if it survives production conditions. For high-throughput inference workloads where latency-per-request matters less than tokens-per-second, this is the kind of number that changes cost models.
• Open weights. You can run it, audit it, and build on it. The HuggingFace release is a base model — not instruct-tuned — which is actually more useful for researchers who want to understand the diffusion behavior.

What deserves a side-eye:

• "98.7% of aggregate benchmark quality" is doing a lot of work. Aggregate means averaged across benchmarks NVIDIA selected. The paper (arXiv 2606.26493) has the methodology — read it before making production decisions. Which benchmarks, what weighting, whether instruction following and reasoning tasks were included.
• The 2.42× throughput measurement was taken under specific hardware conditions. Real-world inference throughput varies substantially by batch size, sequence length, quantization, and hardware generation. This number needs reproduction on your actual stack before you trust it.
• NVIDIA Nemotron Open Model License is not Apache or MIT. It restricts certain uses, particularly using model outputs to train competing models. Worth reading before you build anything commercial on top of this.
• Base model only — no instruction-tuned variant in the July 2 release. Using this in production requires your own fine-tuning investment.

What builders need to know

• The weights are on HuggingFace now: nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16. Base model, BF16. If you have the hardware, you can benchmark it against your own workloads today.
• Benchmark the 2.42× throughput claim against your actual batch sizes and sequence lengths before trusting it. The paper's measurement conditions may not match your inference setup.
• Read the NVIDIA Nemotron Open Model License on the HuggingFace model card before commercial deployment. It is not permissive in the standard open-source sense.
• No instruct-tuned variant is in the July 2 release. Building production use cases requires supervised fine-tuning on your own data.
• The preceding Nemotron-Labs-Diffusion tri-mode paper from May 2026 is worth reading for context on how NVIDIA arrived at the two-tower design.
• Hardware: the model runs at ~3B active parameters per tower. Both towers run during inference. Plan for memory accordingly.

Further reading

• arXiv 2606.26493 — Nemotron-TwoTower paper — architecture, training setup, benchmark methodology; primary source for all performance claims
• HuggingFace model card — nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16 — weights, license, quickstart
• NVIDIA Research — Nemotron-Labs-Diffusion tri-mode (May 2026) — the preceding paper establishing the AR-diffusion hybrid direction
• MarkTechPost — TwoTower overview — accessible summary of the release