Google DeepMind releases DiffusionGemma to speed up local text generation by 4x.

Google DeepMind has released DiffusionGemma, an open AI model designed to make local text generation significantly faster, with Ars Technica reporting that it delivers a 4x speed boost. The launch extends diffusion-based AI beyond its best-known use in image generation and into language tasks, where speed and efficiency matter for running models on devices without relying on remote servers.

According to Ars Technica, the key idea behind DiffusionGemma is to apply diffusion techniques to text output, a method more commonly associated with image synthesis. That matters because faster local inference can make AI assistants, writing tools, and other on-device applications feel more responsive while reducing dependence on cloud infrastructure.