Google DeepMind Converts Gemma 4 Into a Diffusion Model, Hits 1,500 Tokens/Sec

Google DeepMind has published a technical report detailing DiffusionGemma, a text diffusion model created by converting the existing Gemma-4-26B-A4B model rather than training a new architecture from scratch. The model was first released in mid-June 2026; the new report explains the training process and documents where the approach falls short.

Retrofitting instead of retraining

Unlike standard autoregressive language models that generate text one token at a time, DiffusionGemma refines blocks of 256 tokens in parallel, a process similar to how image diffusion models pull a picture out of noise. On an Nvidia H100 accelerator, the model reaches about 1,500 tokens per second, according to Google.

The key claim in the report is that building a diffusion model doesn't require training from zero. Google's team started with the finished Gemma-4-26B-A4B checkpoint and converted it into a diffusion model using less than 10% of the original training token budget, according to the report. Google says DiffusionGemma delivers several times the output speed of both the base Gemma 4 models and prior diffusion models while maintaining comparable accuracy on many tasks.

Two-stage training process

The conversion happens in two stages. First, the model learns to reconstruct noisy text blocks from example data. Second, Google applies a combined phase it calls SD·RL — sampler distillation plus reinforcement learning — merged into a single training process. Reinforcement learning typically improves answer quality, while sampler distillation reduces the number of compute steps needed per output.

According to the report, this combined stage raised quality on reasoning benchmarks by an average of 10 points while nearly quadrupling tokens processed per compute step. A side effect: DiffusionGemma's answers run about 50% shorter than the base model's, which further increases effective speed.

Self-correction through bidirectional generation

Because DiffusionGemma generates reasoning and answer tokens in parallel rather than sequentially, it can revise an early mistake before finalizing output. The report cites a math problem where standard Gemma 4 commits to "-1" as its first output token, then corrects to "-25" later in its derivation — a correction that has to be tacked on after the fact in an autoregressive model. DiffusionGemma can fix such errors during later denoising steps instead.

This bidirectional property also helps with tasks like Sudoku, where each entry depends on values that come later in the sequence. After minimal fine-tuning, DiffusionGemma solves close to 85% of puzzles correctly; Google says the autoregressive base model fails at the task entirely. Structured outputs such as JSON or code repairs reportedly finish in just two to three refinement steps, since the input already constrains most of the output tokens.

Limitations: quality gaps and concurrency

DiffusionGemma trails the autoregressive Gemma 4 on general quality benchmarks despite its speed advantage. Google attributes this to the model being retrofitted rather than trained as a diffusion model from the outset, a relatively short subsequent training phase, and an SD·RL stage tuned to prioritize speed over peak accuracy. The architecture and training data were also carried over unchanged from Gemma 4, which the report notes aren't necessarily optimal for a diffusion setup.

The model can get stuck in repetition loops, occasionally producing the same word multiple times in a row — an artifact of aggressively reduced compute steps. On multimodal tasks, DiffusionGemma sometimes fails to properly close its reasoning section, which the report says artificially deflates benchmark scores.

The throughput advantage also mostly applies to single-user scenarios. Once concurrent requests reach about 32, standard autoregressive models catch up on total throughput, according to Google.

Google DeepMind describes DiffusionGemma as experimental, intended to accelerate research on text diffusion and give developers a base for resource-efficient, specialized adaptations. The model is already used by startup Interfaze for multilingual speech recognition and in a research project on interactive radiology report generation. It's available under an Apache 2.0 license on Hugging Face. Its predecessor, Gemini Diffusion, was demoed by Google in May 2025.

What this means

The practical significance here isn't the model itself — DiffusionGemma trails its autoregressive source on quality — but the conversion method. If a lab can turn an existing production model into a diffusion variant for under 10% of the original training cost, diffusion-based text generation becomes viable to explore without the capital outlay of training a new foundation model. That lowers the barrier for researchers experimenting with parallel generation and self-correcting inference, even if the current tradeoffs — repetition artifacts, weaker multi-user throughput, and lower peak quality — mean DiffusionGemma isn't ready to replace standard LLM serving stacks yet.