NVIDIA introduces Nemotron-Labs-Diffusion, a tri-mode language model that unifies autoregressive (AR), diffusion, and self-speculation decoding within a single architecture. Trained with a joint AR-diffusion objective, the model can switch modes to sustain high throughput across different deployment settings and concurrency levels.
- The study shows AR and diffusion objectives are complementary, with diffusion improving lookahead planning and AR providing left-to-right linguistic priors.
- In self-speculation mode, diffusion drafts while AR verifies, outperforming multi-token prediction methods in acceptance rate and real-device efficiency.
- A speed-of-light analysis demonstrates diffusion's potential, allowing up to 76.5% more tokens per forward pass than self-speculation under an optimal sampler.
- The Nemotron-Labs-Diffusion family scales to 3B, 8B, and 14B parameters, including base, instruct, and vision-language models that outperform state-of-the-art open-source LMs in accuracy and speed.
- Nemotron-Labs-Diffusion-8B decodes 6x more tokens per forward than Qwen3-8B with comparable accuracy, translating to 4x higher throughput on SPEED-Bench with SGLang on a GB200 GPU.
The model consistently outperforms state-of-the-art open-source AR and diffusion LMs in both accuracy and speed, offering significant efficiency gains for deployment.