Google DeepMind releases DiffusionGemma, a model that runs local AI 4x faster

Google DeepMind announced the release of DiffusionGemma, a new member of the Gemma 4 open model family that generates text in parallel rather than sequentially. The model uses a diffusion approach similar to image generation, starting with placeholder tokens and refining them over multiple passes to produce a final block of text. Google said this method makes the model faster and more efficient on local hardware. DiffusionGemma is a Mixture of Experts model with 26 billion total parameters, though only 3.8 billion are activated during inference, allowing it to fit within the 18GB RAM of a high-end GPU. In testing with an Nvidia RTX 5090, the model produced around 700 tokens per second. With a single Nvidia H100 accelerator, it achieved 1,000-plus tokens per second, which Google said is about four times the output of similarly sized autoregressive Gemma models. The approach shifts the bottleneck from memory bandwidth to compute, generating up to 256 tokens in parallel. Google said this offers a measurable boost in non-linear tasks such as in-line editing, molecular sequencing and mathematical graphing. The model was tuned to solve Sudoku puzzles, demonstrating an ability to continuously self-correct large sets of tokens.