Google DeepMind retrofits Gemma 4 into DiffusionGemma

Google DeepMind released DiffusionGemma as a model in mid-June and has now followed up with the technical report behind it. Instead of training a new diffusion model from scratch, the team converted the existing Gemma-4-26B-A4B into a diffusion model using less than ten percent of the original training token budget.

The architecture is the core change. Standard language models generate text one token at a time. DiffusionGemma instead refines blocks of 256 tokens in parallel, similar to how image models pull a picture out of noise, and reaches about 1,500 tokens per second on an Nvidia H100 accelerator.

Getting there took two training stages. First the model learns to reconstruct noisy text blocks from example data. A second, combined phase mixes reinforcement learning with sampler distillation, a method Google calls SD·RL: reinforcement learning typically improves answer quality, while sampler distillation lets the model use fewer compute steps, and Google merges the two into one process. According to the report, this raises quality on reasoning benchmarks by an average of ten points while nearly quadrupling the number of tokens produced per compute step, and as a side effect DiffusionGemma's answers come out about 50 percent shorter, which further boosts speed.

The parallel structure also lets the model correct itself mid-answer. A standard model has to commit to the first digit of a result before it has worked through the reasoning; in a math example from the report, Gemma 4 opens with "-1," works out during its derivation that "-25" is correct, and tacks the correction on afterward. DiffusionGemma develops the answer and the reasoning in parallel, so it can fix an entry before the output is finalized. The same property helps with Sudoku, where every entry depends on entries filled in later: after minimal fine-tuning DiffusionGemma solves close to 85 percent of puzzles correctly, a task the base autoregressive model fails at entirely. Structured outputs such as JSON or code repairs finish after just two to three refinement steps, since the input already fixes most of the tokens. The model keeps its original ability to generate text word by word too, so users can switch between the two modes depending on the task.

Absolute quality still falls short of the autoregressive base model, and Google gives several reasons: DiffusionGemma was retrofitted rather than trained as a diffusion model from the start, the subsequent training phase was relatively short, the SD·RL step prioritized speed over peak quality, and the architecture, training data and other settings were carried over from Gemma 4 rather than chosen for diffusion. The model occasionally falls into repetition loops, producing the same word several times in a row, which the report attributes to the aggressively reduced compute steps. On multimodal tasks it sometimes fails to properly close its reasoning section, which artificially drags down benchmark scores. The speed advantage also mostly holds for a single user at a time; once about 32 concurrent requests hit the model, standard language models catch up on throughput.

Google calls DiffusionGemma explicitly experimental and says the release is meant to speed up research on text diffusion and give the community a foundation for specialized, resource-efficient adaptations. It is already in use: the startup Interfaze applies it to multilingual speech recognition, and a research project is using it for interactive radiology report generation. The model is available under an Apache 2.0 license on Hugging Face. Its predecessor is Gemini Diffusion, which Google demoed in May 2025.

Key facts

• DiffusionGemma refines blocks of 256 tokens in parallel instead of generating one token at a time, reaching about 1,500 tokens per second on an Nvidia H100.
• It was built by converting the existing Gemma-4-26B-A4B into a diffusion model using less than ten percent of the original training token budget, rather than training from scratch.
• A combined RL-and-sampler-distillation step Google calls SD·RL raises reasoning-benchmark quality by an average of ten points, nearly quadruples tokens produced per compute step, and shortens answers by about 50 percent.
• After minimal fine-tuning it solves close to 85 percent of Sudoku puzzles correctly, a task the base autoregressive model fails at entirely, because it can revise entries that depend on ones filled in later.
• The speed edge mostly holds for a single user; once about 32 concurrent requests hit the model, standard autoregressive models catch up on throughput.

Why it matters

DiffusionGemma shows an existing autoregressive model can be retrofitted into a diffusion architecture for under ten percent of the training cost of building one from scratch. That matters because the diffusion approach generates whole 256-token blocks in parallel rather than one token at a time, which both speeds generation up to about 1,500 tokens per second on an Nvidia H100 and lets the model revise its own output before committing to it, something a standard left-to-right model structurally cannot do.

Who it affects

Researchers working on text diffusion get a documented, working retrofit recipe and a released model to build on. Developers who need fast structured output benefit directly: JSON and code repairs finish in just two to three refinement steps. The model is already deployed outside Google, by the startup Interfaze for multilingual speech recognition and in a research project on interactive radiology report generation.

How to use it

DiffusionGemma is available under an Apache 2.0 license on Hugging Face, and it can run in either mode: the parallel diffusion mode for speed, or its original word-by-word autoregressive mode, switchable depending on the task. No pricing or paid tiers are mentioned; the report gives no availability details beyond the open license.

How solid is it

The claims come from Google DeepMind's own technical report, with concrete, checkable results rather than vague promises: a ten-point average lift on reasoning benchmarks, nearly quadrupled tokens per compute step, a jump from total failure to close to 85 percent accuracy on Sudoku after minimal fine-tuning, and two-to-three-step convergence on structured output. Google itself labels the model experimental rather than production-ready.

Risks and caveats

By Google's own account, DiffusionGemma's absolute performance still trails the autoregressive base model, for several stated reasons: it was retrofitted rather than trained as a diffusion model from the start, the extra training phase was short, the SD·RL step traded peak quality for speed, and the underlying architecture and training data were inherited from Gemma 4 rather than optimized for diffusion. It occasionally gets stuck repeating individual words, an artifact of its reduced compute steps, and on multimodal tasks it sometimes fails to close its reasoning section properly, which artificially lowers its benchmark scores. Its throughput advantage over standard models also disappears once concurrency reaches around 32 simultaneous requests.