Google DeepMind’s DiffusionGemma points to a practical path for text diffusion: adapt an existing language model instead of starting over. The model begins from Gemma-4-26B-A4B and changes how text is produced, shifting from token-by-token generation to parallel refinement of larger text blocks.
The result is not presented as a finished replacement for standard language models. It is described as an experimental model, and the technical report lays out both the benefits and the compromises that come with this approach.
A different way to generate text
Standard language models produce text one token at a time. DiffusionGemma can refine blocks of 256 tokens in parallel, a process the source compares to the way image AI systems turn noise into a finished picture.
That difference matters because it changes where the model can gain speed. On an Nvidia H100 accelerator, DiffusionGemma reaches about 1,500 tokens per second. The model also keeps its original ability to generate text word by word, so users can choose between generation modes depending on the task.
The larger point is that Google DeepMind did not build a new model from zero. The team converted Gemma-4-26B-A4B into a diffusion model using less than ten percent of the original training token budget, according to the report. That makes DiffusionGemma important not only as a model, but also as a demonstration of a more resource-conscious adaptation strategy.
How the training process balances quality and speed
The conversion uses two main training stages. In the first stage, the model learns to rebuild noisy blocks of text from example data. This gives it the foundation for diffusion-style generation, where the answer is refined across a block instead of written strictly from left to right.
The second stage combines reinforcement learning with sampler distillation. Google calls this combined process SD·RL. In plain terms, the report presents it as a way to improve answer quality while also reducing the number of compute steps the model needs.
According to the report, SD·RL improves quality on reasoning benchmarks by an average of ten points while nearly quadrupling the number of tokens produced per compute step. It also makes DiffusionGemma’s answers about 50 percent shorter, which adds to the speed gain.
Those shorter answers are part of the model’s speed profile, but they also show the tradeoff at the center of the system. DiffusionGemma is tuned to move quickly, and the report is clear that this choice affects peak performance.
Why bidirectional reasoning is useful
One of DiffusionGemma’s most interesting advantages is that it can develop different parts of an answer together. A standard language model may commit to the first part of an answer before the later reasoning has fully played out. DiffusionGemma can refine the answer and the reasoning in parallel, giving it a chance to correct mistakes before the final output is set.
The report includes a math example where Gemma 4 begins with -1, then reaches -25 during its derivation and adds a correction later. DiffusionGemma’s process is better suited to that kind of situation because the model is not forced to lock in the early output before later steps are resolved.
Sudoku shows a related strength. Each entry can depend on entries that appear later, so a strictly left-to-right process is not always a natural fit. After minimal fine-tuning, DiffusionGemma solves close to 85 percent of puzzles correctly, while the base model fails at the task entirely.
Structured outputs also benefit from the same logic. The report says tasks such as JSON generation or code repairs can finish after just two to three refinement steps, because much of the output is already constrained by the input.
The limits are still important
DiffusionGemma does not beat the autoregressive base model in absolute performance. Google points to several reasons. It was retrofitted rather than trained as a diffusion model from the beginning, and the follow-up training phase was relatively short.
The second stage, SD·RL, was also designed with speed in mind rather than maximum quality. The architecture, training data and other settings were carried over from Gemma 4, even though they may not be ideal for a diffusion model.
The report also describes practical issues in generation. DiffusionGemma can get caught repeating individual words multiple times in a row, which is linked to the sharply reduced number of compute steps. On multimodal tasks, it can fail to close its reasoning section correctly, which hurts benchmark results.
The speed advantage has a deployment caveat as well. It is strongest in single-user scenarios. Once about 32 concurrent requests reach the model, standard language models catch up on throughput.
Why the release matters for AI research
Google explicitly frames DiffusionGemma as experimental. The release is meant to accelerate research into text diffusion and provide a base for specialized, resource-efficient adaptations.
That framing is important. DiffusionGemma is not simply a faster text generator; it is a test of whether existing models can be redirected into a different generation style without repeating the full cost of original training. The answer from this report is that the approach can work, but it carries visible tradeoffs.
The model is already being used by the startup Interfaze for multilingual speech recognition and in a research project on interactive radiology report generation. Google previously made it available under an Apache 2.0 license on Hugging Face. Its predecessor is Gemini Diffusion, which Google demoed in May 2025.