DiffusionGemma, Google’s open model that writes text in parallel
DiffusionGemma
Every chatbot you have used writes its answer one word at a time. Google DeepMind’s DiffusionGemma writes a whole block of words at once, and sharpens them together. It is an open Gemma model, one you can download and run yourself, and the way it writes is borrowed from AI image generators.
About 1,500 word-pieces a second on one chip, at a real accuracy cost
DiffusionGemma is an open model from Google DeepMind, under Apache-2.0, so anyone can download and build on it. The weights have been out since June. What is new this week is the technical report that explains how the model works. The report is trending today. On one data-center graphics chip, answering a single request, Google reports about 1,500 word-pieces a second. A word-piece is just a small chunk of text. That is about seven times the ordinary Gemma it was built from. But it is less accurate than that same Gemma, and that trade is the whole story.
A whole block at once, instead of one word at a time
What is genuinely new is how it writes. Every chatbot works the same way, one word at a time. The model predicts a word, feeds it back, predicts the next, and the sentence grows only left to right. That is the speed bottleneck. DiffusionGemma borrows the idea behind AI image generators. An image generator starts from a rough, noisy picture and sharpens it over a few passes. DiffusionGemma does the same thing with words. Google calls it discrete diffusion decoding.
About 20 word-pieces filled in on each pass, across a 256-word canvas
256 blanked
about 20 filled
block done
dozens of trips
for a paragraph
DiffusionGemma starts each block as a canvas of 256 word-pieces, all blanked out. Then, in one trip through the model, it looks at the whole canvas and fills in its best guess for many positions at once. It repeats that a few times, each pass sharpening the block. Google reports it resolves about 20 word-pieces on every pass, and finishes a block in around a dozen passes. So a paragraph that the one-word way needs dozens of separate trips for takes only a handful. That is where the speed comes from: far fewer trips, the sevenfold jump on one chip from before. The blocks are still produced in order; only the words inside each block are filled in together.
Built by converting Gemma 4, on under a tenth of its training data
Google did not train this from scratch. They took the ordinary Gemma 4 and converted its text half to work by diffusion, in two stages, on under 10% of the data the original Gemma needed. Stage one taught it to fill in blanked-out words using the words on both sides. An ordinary chatbot can never do that, because it only sees what came before. Stage two polished accuracy and speed together, squeezing many sharpening passes down to a few. Because it was converted, not rebuilt, it keeps the parent’s skills. It still reads images and video, and can even fall back to ordinary one-word writing with only minor loss.
Behind its parent across the board, closest on knowledge and furthest on math
These numbers are all Google’s own reported tests, DiffusionGemma against the ordinary Gemma 4 it was built from. It trails the parent across the board, and the gap is uneven. On broad knowledge it is nearly level, 77.6 to the parent’s 82.6. On graduate-level science questions, 73.2 to 82.3. On coding problems, 69.1 to 77.1. On hard competition math the gap blows open, 69.1 to 88.3. It holds up on broad knowledge, and falls furthest behind on the hardest math. Google’s report frames the result that way itself, a new point on the trade between speed and accuracy, not a win.
Open weights on Hugging Face, the report on arXiv
Everything is out in the open. The weights are on Hugging Face, ungated, under the same permissive Apache-2.0 license. And the technical report, this week’s news, is on arXiv. The links are on screen.
The speed is real, and so is the cost in accuracy
A chatbot writes one word at a time. DiffusionGemma borrows the trick behind image generators to sharpen a whole block at once, and runs several times faster on a single chip. Google’s own numbers show that speed costs accuracy, most on the hardest reasoning. The interesting part is not a leaderboard. It is a genuinely different way to write text, and the first clear map of what it costs.
