NVIDIA releases Nemotron-Labs Diffusion series models with three generation modes
Original title:Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models
AI summary
NVIDIA released the Nemotron-Labs Diffusion series, including 3B, 8B, and 14B text models and an 8B vision-language model (VLM), all under the commercially friendly NVIDIA Nemotron Open Model License or NVIDIA Source Code License. The models support three modes: autoregressive, diffusion (parallel block-by-block generation followed by progressive refinement), and self-speculative (diffusion drafts candidate tokens, then autoregressive verification). The 8B model's average accuracy is 1.2% higher than Qwen3 8B; tokens per forward pass (TPF) in diffusion mode reach 2.6x that of autoregressive, and self-speculative reaches 6–6.4x. The models were pretrained on 1.3T tokens and fine-tuned on 45B tokens. Code and models have been released on HuggingFace and GitHub, and inference will be supported by SGLang.
Key Takeaways
- NVIDIA released the Nemotron-Labs Diffusion series,including 3B,8B,and 14B text models and an 8B…
- The models support three modes: autoregressive,diffusion (parallel block-by-block generation…
- The 8B model's average accuracy is 1.
- 2% higher than Qwen3 8B
- tokens per forward pass (TPF) in diffusion mode reach 2.
Quick Facts
- Published
- Updated
- Source
- Hugging Face 社区博客(混合发现)
- Category
- Model Releases
- AI score
- 63
自推测模式让文本生成速度飙到AR模型的4倍,而且输出质量无损。NVIDIA这次开源的不仅是个新模型,更是一套能直接用在现有流程里的加速方案。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: NVIDIA releases Nemotron-Labs Diffusion series models with three generation modes. May 23, 2026. https://new.nat6.net/en/news/cmphlmb8l0peosljw5rpq8dt2/