Diffusion Models Cross into Text Generation: Nvidia Open-sources Nemotron-Sigma, Decoding Speed Up to 4.2x Faster Than Autoregressive
NVIDIA has released the continuous diffusion language model Nemotron-Labs-Sigma and open-sourced 14 weight checkpoints with parameter sizes of 3B and 8B. This model adapts the diffusion architecture, originally used primarily in image generation, to text generation. It performs denoising generation in the trained 16-dimensional embedding space and supports few-step sampling distillation, with a decoding speed up to 4.20 times faster than traditional autoregressive decoding, offering a new exploration direction for text generation technology routes.
The Nvidia Nemotron Labs team has recently launched the new Nemotron-Labs-Sigma continuous diffusion language model, and simultaneously released 14 open-source weight checkpoints, covering two parameter sizes: 3B and 8B.
Currently, most commercial and open-source large language models (LLMs) adopt an autoregressive architecture. When generating text, tokens are output one by one in sequence, and decoding speed decreases linearly with text length. Anyone who has run local LLMs will agree that for slightly larger parameter versions, the waiting time is very noticeable when generating long texts. In scenarios with high concurrency and edge deployment, the computational cost and latency of this type of architecture are difficult to reduce.
Sigma adapts the generation logic of diffusion models to the text domain. Previously, diffusion architectures were mainly used in image and video generation, with the core logic of gradually removing noise from noisy random data to restore valid content that meets requirements. This model first maps text to the trained 16-dimensional embedding space, performs denoising generation in the continuous space, and then decodes it into readable text. It also supports distillation into a few-step sampler to further compress the generation process.
The officially released test data shows that the optimized Sigma’s decoding speed is up to 4.20 times faster than traditional autoregressive decoding.
The 14 checkpoints currently released cover different training stages and sampling configurations, and developers can directly pull the weights for adaptation testing. The diffusion-based approach for text generation is still a relatively niche technical direction. Besides the speed advantage, core metrics such as semantic coherence and long-text accuracy still require more real-world testing for verification.
Information source: MarkTechPost, author Asif Razzaq
发布时间: 2026-11-20 22:44