DiffusionGemma: an open-source diffusion model that generates text 4x faster
Original title:DiffusionGemma: 4x faster text generation
AI summary
Google DeepMind released DiffusionGemma, an open-source experimental model that uses text diffusion to break away from autoregressive token-by-token generation, producing 256 tokens in parallel per forward pass. The 26B MoE model activates only 3.8B parameters at inference and fits on 18GB consumer GPUs after quantization. It reaches 1000+ tokens/s on H100 and 700+ tokens/s on RTX 5090, a 4x speedup. With bidirectional attention and self-correction, it targets local interactive workflows such as inline editing and code infilling, and is released under the Apache 2.0 license.
Key Takeaways
- Google DeepMind released DiffusionGemma,an open-source experimental model that uses text diffusion…
- The 26B MoE model activates only 3.
- 8B parameters at inference and fits on 18GB consumer GPUs after quantization.
- It reaches 1000+ tokens/s on H100 and 700+ tokens/s on RTX 5090,a 4x speedup.
- With bidirectional attention and self-correction,it targets local interactive workflows such as…
Quick Facts
- Published
- Updated
- Source
- Google DeepMind:Blog(RSS)
- Category
- Model Releases
- AI score
- 72
DiffusionGemma 虽为实验性质,但它把文本生成从“串行打字机”变成了“并行印刷机”,本地推理速度 4 倍提升,对需要实时交互的开发者是个值得关注的方向。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: DiffusionGemma: an open-source diffusion model that generates text 4x faster. June 11, 2026. https://new.nat6.net/en/news/cmq8aoxud00ngslldd191n34b/