DeepMind publishes details about “DiffusionGemma,” a new approach aimed at speeding up text generation. The work is presented as a method that uses diffusion-based modeling paired with the Gemma model family, with the stated goal of reducing generation latency while keeping output quality competitive. In its description, DeepMind claims that DiffusionGemma can generate text faster than prior baselines, including a reported speedup of up to 4x. The release discusses the motivation for diffusion-based techniques in generative tasks and explains how the method is applied to text generation. It also emphasizes evaluation against existing approaches, focusing on generation speed and performance outcomes. Outside of DeepMind’s post, the provided Hacker News material contains only a brief reference to the same topic without additional technical or evaluative details. No other specific quantitative results, benchmarks, or limitations are included in the supplied excerpts beyond the headline claim of significantly improved generation speed.