June 10, 2026
Google has released DiffusionGemma, an open-weight language model that generates text by refining whole blocks of tokens at once rather than one word at a time, and which the company said runs up to four times faster than conventional models on dedicated graphics processors. The weights are available on Hugging Face under an Apache 2.0 licence.
The model departs from the autoregressive design used by most large language models, which predict text one token after the next. According to Google, DiffusionGemma drafts a block of 256 tokens and revises it over several passes, adapting the denoising technique used in image generation. The company reported throughput above 1,000 tokens per second on a single Nvidia H100 and above 700 on a consumer RTX 5090.
DiffusionGemma is a mixture-of-experts model with 26 billion total parameters, of which Google said about 3.8 billion are active at each step. The company said the model fits within 18GB of memory when quantised, within range of high-end consumer graphics cards.
Google was explicit about the cost of the approach. It said output quality is lower than its standard Gemma models, and described DiffusionGemma as experimental, better suited to speed-critical tasks such as in-line editing and code infilling than as a general-purpose model.
For where the standard Gemma line stands on small hardware, our own measurements of Gemma 4 E4B on edge devices are the comparison point: the ordinary decode loop, sped up by multi-token prediction rather than replaced by diffusion.
The release adds to a small body of research on text diffusion, an approach that has drawn interest as a route to faster generation but has not yet matched autoregressive systems on quality at scale. Because the model ships under a permissive licence that allows commercial use, outside researchers can now test those speed and quality claims independently, using the weights and the model card Google has posted on Hugging Face.