Google Turned a Finished Model Into a Much Faster One, and Published the Recipe
DiffusionGemma writes 256 words at once instead of one after another, hits about 1,500 tokens per second, and cost less than ten percent of a fresh training run. The technical report explains the tradeoffs.
Almost every AI chatbot you have used writes the way you would type a text message: one word at a time, left to right, each word chosen after the one before it. Google DeepMind has now published the technical report for DiffusionGemma, a model that works differently. It refines whole blocks of 256 tokens at once, the way an image generator pulls a picture out of noise, and it does so at roughly 1,500 tokens per second on an Nvidia H100 chip. A token, roughly speaking, is a chunk of a word, so that is a very fast typist.
The headline of the report is not the speed, though. It is the cost. Google did not train this from scratch. The team took the existing Gemma-4-26B-A4B model and converted it into a diffusion model using less than ten percent of the original training budget. Two stages did the work: first the model learns to reconstruct deliberately noisy blocks of text, then a combined phase Google calls SD-RL mixes reinforcement learning, which improves answer quality, with sampler distillation, which teaches the model to get by with fewer compute steps. That combination lifted reasoning scores by about ten points on average while nearly quadrupling how many tokens each compute step produces.
There is one genuinely charming side effect. A normal model has to commit to the first digit of an answer before it has finished the reasoning. The report shows Gemma 4 starting a maths answer with “-1,” realising mid-derivation that “-25” is right, and tacking on a correction. DiffusionGemma develops answer and reasoning in parallel, so it can quietly fix the wrong number before anything is final. On Sudoku, where every entry depends on entries that come later, the diffusion version solves close to 85 percent of puzzles after light fine-tuning. The base model fails the task entirely.
Now the caveats, because Google is unusually candid about them. Absolute quality still trails the original autoregressive Gemma 4, especially on reasoning. The model sometimes gets stuck repeating individual words, an artefact of the aggressively trimmed compute steps. On multimodal tasks it occasionally forgets to close its reasoning section, which drags scores down. And the speed advantage largely holds for a single user: once roughly 32 requests hit the model at the same time, conventional models catch up on throughput. Google calls it an experimental release, published under Apache 2.0 on Hugging Face so others can build on it.
What this means for you: If you just use a chatbot, nothing changes today. But this is a decent preview of where “fast and cheap” is heading, and speed is the thing most people actually notice. For anyone running models themselves, the interesting claim is the retrofit: you may not need a fresh multi-million-dollar training run to get a fundamentally different model, just a tenth of one applied to a model you already have. A startup is already using DiffusionGemma for multilingual speech recognition, which is roughly the point of putting the weights out in public.
Sources
Source: https://arxiv.org/abs/2608.00146
Fraudsters Are Enrolling Students Who Do Not Exist, Then Letting AI Do the Homework
US community colleges are dealing with a new kind of scam: fake enrolments to collect financial aid, with chatbots quietly completing the coursework that keeps the money flowing.
