Diffusion language model
A model that writes by starting from noise or masked text and refining the whole passage over several steps.
Also: text diffusion
Diffusion models became famous for images. Diffusion language models apply the idea to text: they start with a sequence of masked or random tokens and repeatedly improve all positions at once, rather than writing left to right. This allows many tokens to be produced in parallel, so generation can be much faster. Inception Labs released Mercury in 2025, Google showed an experimental Gemini Diffusion the same year, and research models such as LLaDA explored the approach. Quality has generally trailed the best autoregressive models of similar size.