Breaking

NVIDIA Releases Diffusion LM Nemotron-Labs-TwoTower With 2.42x Faster Generation

July 1, 2026 at 15:00 EDT

Foundation Models
Open Source
Research & Papers

On July 1, 2026, NVIDIA Research released Nemotron-Labs-TwoTower, a diffusion language model that splits an existing 30B-scale model into two towers to generate tokens in parallel. The open-weight model reportedly retains 98.7% of the base model's quality while boosting generation throughput by 2.42x.

July 1, 2026 · NVIDIA Research

TwoTower: A Diffusion LLM That Generates 2.42× Faster

NVIDIA splits its autoregressive Nemotron-3-Nano into two towers — a frozen Context Tower and a trainable Denoiser Tower — generating tokens in parallel to boost throughput while preserving 98.7% of quality.

2.42×

Generation throughput vs AR baseline

98.7%

Aggregate quality retained

~60B

Total params · ~3B active per token

Throughput, drawn to scale

Parallel block generation replaces one-token-at-a-time decoding.

1.0×

AR baseline

2.42×

TwoTower

Two-Tower Architecture

Context Tower

Frozen · processes clean tokens causally

→

Denoiser Tower

Trainable · refines noisy blocks

Linked by layer-aligned cross-attention · trained on ~2.1T tokens · one checkpoint switches Diffusion / Mock-AR / AR modes

Benchmark: AR baseline vs TwoTower

MMLU (5-shot)78.56 → 78.24

HumanEval79.27 → 75.58

GSM8K92.49 → 90.14

Bars show TwoTower as a share of AR baseline — the gap is most visible on code and math.

The upside

Breaks AR's sequential bottleneck
Multi-mode support from one checkpoint
Practical layer-aligned cross-attention design

The caveats

Slight quality drop to 98.7%
Two GPUs recommended for inference
No large-scale production benchmarks yet

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.