# Google released DiffusionGemma to test parallel text generation > The experimental open model generates blocks through diffusion rather than token by token, targeting interactive local workloads on GPUs. Canonical URL: https://www.devobs.io/news/news-diffusiongemma-experimental-text-generation/ By: Sofia Reyes Published: 2026-09-06T11:58:54.627Z Updated: 2026-09-06T11:58:54.627Z Event date: 2026-06-10 Section: AI Google introduced DiffusionGemma, an experimental open model that generates and refines blocks of text in parallel instead of decoding strictly from left to right. The [June 10 release](https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/) describes a 26-billion-parameter mixture-of-experts model that activates 3.8 billion parameters during inference and is released under Apache 2.0. Google targets local, speed-sensitive workflows on dedicated GPUs and explicitly says standard autoregressive Gemma 4 remains preferable when maximum output quality matters. ## Parallel decoding changes the performance envelope The architecture can use accelerator compute differently from a conventional token stream, especially at low or moderate batch sizes. Google also warns that the advantage can shrink in high-throughput serving, where autoregressive systems batch many requests efficiently. Teams should therefore resist turning the publisher's reported peak speedups into a general serving claim. Block refinement also changes application behavior. Streaming interfaces, cancellation, token accounting and structured-output validation may behave differently when a whole span is revised across iterations. ## Evaluate quality and latency together A useful test should run on the exact GPU, quantization and inference stack planned for deployment. Compare time to useful output, final latency, power, memory and task success against an autoregressive baseline. Include code infilling and editing, where bidirectional context may be useful, alongside ordinary generation where quality tradeoffs may dominate. Because Google labels the model experimental, isolate it behind a versioned interface and retain a fallback. Validate licenses for the complete distribution stack, not only the model weights. The release is best read as an architecture experiment that developers can measure, rather than a drop-in replacement for every local language model. ## Source references -