# Google’s DiffusionGemma Cuts Training Cost for Faster Text Generation

> Source: <https://mlq.ai/news/googles-diffusiongemma-cuts-training-cost-for-faster-text-generation/>
> Published: 2026-08-09 19:35:45.383447+00:00

# Google’s DiffusionGemma Cuts Training Cost for Faster Text Generation

- Google adapted Gemma 4 into DiffusionGemma using less than 10% of the starting model’s training-token budget.
[[1]](https://arxiv.org/abs/2608.00146) - The model refines 256-token blocks in parallel and reaches about 1,500 tokens per second on one NVIDIA H100, according to Google’s technical report.
[[1]](https://arxiv.org/abs/2608.00146) - DiffusionGemma remains behind Gemma 4 on several reasoning benchmarks, while independent tests suggest its practical advantage is fast, bidirectional editing rather than a broad quality win.
[[2]](https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/)[[3]](https://ai.google.dev/gemma/docs/diffusiongemma/model_card)

Google DeepMind has published a technical report on DiffusionGemma, an experimental open-weight model that converts the Gemma 4 architecture into a discrete text-diffusion system using less than 10% of Gemma 4’s original training-token budget. [[1]](https://arxiv.org/abs/2608.00146)

## A cheaper route to parallel decoding

DiffusionGemma starts with a 256-token canvas and repeatedly refines it, allowing tokens within the block to attend to one another instead of being generated strictly from left to right. The 26-billion-parameter mixture-of-experts model activates 3.8 billion parameters during inference and reached roughly 1,500 output tokens per second on a single NVIDIA H100 in Google’s evaluation. [[1]](https://arxiv.org/abs/2608.00146)

Google released the model under the Apache 2.0 license on June 10, 2026, while warning that its overall output quality is lower than standard Gemma 4. The company positions it for speed-sensitive local workflows such as inline editing, code infilling and other tasks where future and preceding text need to be considered together. [[2]](https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/)

## Speed does not erase the quality gap

The model card reports lower scores than the matched Gemma 4 26B model on MMLU Pro, AIME 2026, LiveCodeBench, GPQA Diamond and several multimodal evaluations. For example, DiffusionGemma scored 69.1% on AIME 2026 without tools versus 88.3% for Gemma 4, and 73.2% on GPQA Diamond versus 82.3%. [[3]](https://ai.google.dev/gemma/docs/diffusiongemma/model_card)

Independent researchers reached a more limited conclusion: in a 686-prompt study, DiffusionGemma’s accuracy was comparable to its autoregressive sibling on the tested scorable tasks, but its decoding was only partly parallel and showed a measurable left-to-right bias. [4] That suggests the release’s practical significance is narrower than a wholesale replacement for autoregressive language models: it provides an open platform for fast local generation, infilling and task-specific fine-tuning while leaving general reasoning performance unresolved.

## Companies mentioned

## Further sources

[[1] Google DeepMind’s DiffusionGemma Technical Report states that the model is fine… ↗](https://arxiv.org/abs/2608.00146)

[[2] Google’s June 10, 2026 launch post describes DiffusionGemma’s Apache 2.0 licens… ↗](https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/)

[[3] Google’s DiffusionGemma model card lists benchmark comparisons with Gemma 4 26B… ↗](https://ai.google.dev/gemma/docs/diffusiongemma/model_card)

[[4] An independent arXiv study by Ali Asaria, Tony Salomone and Deep Gandhi evaluat… ↗](https://arxiv.org/abs/2606.14620)

The stories that matter, in one email. Free — unsubscribe anytime.