Google's DiffusionGemma AI Achieves 1,000 Tokens Per Second

On June 10, 2026, Google introduced DiffusionGemma, an open-weight AI model that operates using a text diffusion method to create blocks of text. This allows it to achieve speeds exceeding 1,000 tokens per second on an NVIDIA H100 graphics card, which is four times faster than conventional autoregressive models like its predecessor, Gemma 4.

DiffusionGemma generates entire blocks of 256 tokens in parallel, a significant shift from the traditional method of generating text word-by-word, which constrains each token to be dependent on its predecessor. This innovative approach enables bidirectional attention, where every token can reference all others during creation, enhancing performance in tasks where the conclusion informs the start of the response. However, Google acknowledges that while this new model excels in speed, it does not match the output quality of Gemma 4.

Despite its capabilities, using DiffusionGemma requires specific configurations that are not yet widely available in public runtimes. The requisite custom drafter module for local inference is absent from platforms such as mlx-lm or LM Studio, limiting its usability for general consumers. As noted by Google in their announcements, although DiffusionGemma promises rapid generation speeds, its practical deployment remains constrained by these technical barriers.

The demand for high-speed text generation tools continues to grow, particularly for applications in AI agents and autonomous workflows. Observers will be interested in how Google approaches the issue of access and usability for the broader developer community, as well as the impact this model might have on the evolving landscape of AI-generated text.