DiffusionGemma experiments with generating text in parallel blocks

Google introduced DiffusionGemma on 10 June as an experimental Apache 2.0 model for parallel text generation. The 26B mixture-of-experts model activates 3.8B parameters and drafts 256-token blocks rather than producing every token in a left-to-right sequence. Google reported generation up to four times faster on dedicated GPUs.

The qualification is central to the release. Google says the benefit is strongest for local or lower-concurrency use; busy cloud servers already batch conventional model requests effectively. It also recommends standard Gemma 4 when output quality is the priority, acknowledging a trade-off in the experimental model. DiffusionGemma therefore offers a research route for interactive editing and infilling, rather than a drop-in promise of cheaper, better inference for every service.