Google releases DiffusionGemma, a 4x faster text generation model
DiffusionGemma: 4x Faster Text Generation
Google announced DiffusionGemma on its official blog, a text generation model built with diffusion methods that runs up to 4x faster than similarly sized autoregressive models. It adapts image diffusion techniques to text—generating a full passage in parallel and then refining it through denoising steps. The post claims this cuts latency significantly for real-time use cases. The body doesn't disclose parameter count, benchmark scores, or whether it will be released as open weights or an API.
Why it matters: Google detailed DiffusionGemma, applying image diffusion's parallel generation and denoising to text, with latency a quarter of same-size autoregressive models.