News › 2025 › December 2025
Diffusion LLMs promise faster generation
Research2025-12-18
Source: arxiv.org
Diffusion-based language models show parallel decoding that could beat autoregressive speed.
Autoregressive models emit one token at a time; diffusion language models denoise many at once, and new results suggest the approach can be both fast and competitive on quality.
If it holds up it changes the economics of serving — background for the models on the text & LLMs page.