Google’s DiffusionGemma team has posted the technical report behind the model, and its headline throughput number differs from what the company said when the model shipped. AI Insiders covered the June 12 launch, when Google cited throughput above 1,000 tokens per second on a single H100. The report, submitted to arXiv on July 31, states a higher figure: roughly 1,500 tokens per second on the same chip. The paper does not explain the gap. It simply reports the larger number as its measured result across the authors’ evaluation suite.
DiffusionGemma works by discrete diffusion rather than standard left-to-right decoding. Instead of producing one token after another in sequence, the model begins with a masked 256-token block and reworks the whole block over several passes until it settles on an answer. That shift converts what is normally a sequential bottleneck into a parallel one, which the paper credits as the source of the speed gain.
The model is not a fresh pretraining run. The report says the team built DiffusionGemma by adapting Gemma 4, a mixture-of-experts model with 3.8 billion active parameters out of 25.2 billion total, through a two-stage process that used less than one-tenth of the token budget spent on the original autoregressive version. Stage one taught the model to denoise text in both directions at once through supervised fine-tuning. Stage two paired reinforcement learning with a distillation step aimed at the sampler itself, which the authors say lifted output quality and speed together rather than trading one for the other.
The authors describe this as pushing the Pareto frontier for trading generation speed against model capability, and they credit the 1,500-token figure with beating autoregressive systems even when those systems lean on the fastest available speculative-decoding tricks. What the abstract does not supply is a benchmark table showing how much quality DiffusionGemma gives up against a standard autoregressive model built from the same base. That comparison, not the speed number, is the test a reader should look for once the full paper circulates.
The paper adds that the diffusion version keeps Gemma 4’s thinking-mode reasoning, its multimodal input handling, and its long-context window. It also states that DiffusionGemma can still produce text the old-fashioned way, one token at a time, losing only a small amount of performance when it does, which the authors position as an opening toward models that mix both decoding styles. The abstract does not describe how or when such a hybrid system would be tested.
Anyone benchmarking DiffusionGemma against the June announcement should update their throughput assumption to roughly 1,500 tokens per second on an H100, and hold off on capability comparisons until the report’s full evaluation tables, not just its abstract, are available.
The DiffusionGemma Team at Google published “DiffusionGemma Technical Report” on arXiv (2608.00146), submitted by Jean Tarbouriech on July 31, 2026.
