Google unveils DiffusionGemma for faster text generation

Google has introduced DiffusionGemma, an experimental open-source model designed for significantly faster text generation. Unlike traditional autoregressive models that produce text token by token, DiffusionGemma generates entire blocks of text simultaneously, achieving up to four times the inference speed on dedicated GPUs. This approach is particularly beneficial for speed-critical, interactive local workflows such as in-line editing and rapid iteration. While its overall output quality is lower than standard Gemma models, its bi-directional attention and iterative self-correction capabilities make it suitable for specific tasks and developer exploration.