Nemotron TwoTower 30B A3B Base BF16
nvidia/Nemotron-TwoTower-30B-A3B-Base-BF16
Open Source · chat · open-weights
Open
Alert me on changes
Context
—
Max output
—
Weights
Open
API $/1M
—
Modalities
text
Released
11 Apr 2026
License: other · nvidia/Nemotron-TwoTower-30B-A3B-Base-BF16
AI summary
● machine-written
NVIDIA Nemotron TwoTower 30B reaches 2.42× generation throughput via diffusion
Nemotron TwoTower 30B A3B is a diffusion language model that splits generation into a frozen autoregressive context tower and a trained denoiser tower, enabling parallel token refinement. Built on the Nemotron-3-Nano-30B-A3B backbone with hybrid Mamba-2/attention/MoE architecture, it achieves 2.42× higher wall-clock throughput while retaining 98.7% of autoregressive baseline quality. The model is released as open weights under the NVIDIA Nemotron Open Model License and supports three inference modes: full two-tower diffusion, AR-only, and mock-AR decoding.
What's new
- Two-tower architecture separates context (frozen AR) and denoiser (trained) functions to enable parallel token generation
- Achieves 2.42× higher generation throughput versus autoregressive baseline while maintaining 98.7% benchmark quality
- Requires 2 GPUs with ~59GB per GPU for BF16 full two-tower diffusion, or single 80GB GPU for AR-only mode
- Trained on ~2.1T tokens; backbone pretrained on 25T tokens with 128 routable MoE experts (6 active + 2 shared)
- Supports block-wise decoding with configurable block size (S=16 default) and refinement steps per block
Best for
High-throughput text generation requiring parallel token refinementDeployment on H100/A100 GPU clusters with sufficient VRAMUse cases tolerating modest quality tradeoff (1.3%) for 2.4× faster generationMultilingual and commonsense reasoning tasks (scores match or exceed AR baseline)
Source: https://huggingface.co/nvidia/Nemotron-TwoTower-30B-A3B-Base-BF16