Google releases DiffusionGemma, a diffusion-based text model that generates 4× faster than autoregressive models
Mythos阴影里谷歌悄悄发模型,速度暴涨4倍
Google open-sourced DiffusionGemma, a 26B MoE diffusion text model that activates only 3.8B parameters at inference and fits in 18GB VRAM after quantization. It denoises 256 tokens in parallel—like a printing press instead of a typewriter—hitting 1,000+ tokens/s on an H100 and 700+ on an RTX 5090, roughly 4× faster than a comparable autoregressive model. Bidirectional attention enables real-time self-correction; after fine-tuning, Sudoku accuracy jumped from 0% to 80%. Quality still trails Gemma 4, and Google positions it as an experimental “racehorse” for speed-sensitive local use. Released under Apache 2.0, weights available on Hugging Face.
Why it matters: Google open-sources DiffusionGemma, applying diffusion's denoise-a-whole-passage idea to text generation, hitting 1000+ tokens per second on H100 and 700+ on RTX 5090, about 4x a same-size autoregressive model.