Almost every AI model you've used writes the way you'd read a sentence aloud: one word after another, left to right, each token chosen before the next. Google just open-sourced a model that doesn't. DiffusionGemma, released on 10 June, generates text the way image generators make pictures — and in doing so runs up to four times faster.

The trick: writing like you'd paint

Image models like the ones behind AI art start with a field of random noise and repeatedly denoise it — nudging the whole canvas toward a coherent picture in parallel, a bit at a time, all at once. DiffusionGemma applies that same idea to words. Instead of committing to token one, then token two, then token three in strict sequence, it refines a whole block of text in parallel across several passes.

The payoff is speed. Because it isn't bottlenecked by that one-word-at-a-time chain, DiffusionGemma hits 700+ tokens per second on a single high-end consumer GPU (an RTX 5090), and over 1,000 per second on a data-centre H100 — roughly four times faster than the equivalent conventional model. Nvidia has already optimised it to run locally on its chips.

The grown-up caveat

Here's the part the breathless posts skip: speed costs quality, for now. DiffusionGemma scores lower than the standard, autoregressive Gemma 4 on the established benchmarks, and Google is refreshingly blunt about it — the model ships labelled experimental, with an explicit recommendation to keep using Gemma 4 for production work where output quality matters more than raw speed.

So this isn't "Google's new best model." It's a research bet, put in the open for people to poke at, that a fundamentally different way of generating text might one day be both fast and good. The specs underline the "research" framing: it's a mixture-of-experts model with 25.2 billion total parameters but only 3.8 billion active at once, and it's released under a permissive Apache 2.0 licence — free to use, modify and build on.

Why it matters

Text diffusion has been a promising lab idea for a while; a major lab shipping a usable, open, GPU-accelerated version of it is the news. If the quality gap closes, "generate a whole answer in a few parallel passes" could reshape the economics of running AI — the same cost-of-inference pressure we've been tracking all week, attacked from the architecture side rather than the billing side.

For now, the honest summary is the one Google itself gives: brilliantly fast, not yet as sharp, and out in the open for everyone to improve. That's exactly the kind of bet worth watching — and exactly the kind that's easy to over-hype. We'll take the fast-but-experimental framing over the hype.