Google's DiffusionGemma proves you don't need to train from scratch to build a text diffusion model

DiffusionGemma is a text diffusion model retrofitted from Gemma using minimal training compute. Google DeepMind adapted the existing architecture to generate 256 tokens in parallel, achieving speeds of roughly 1,500 tokens per second. While generation speed increases significantly, output quality still lags behind the original autoregressive model, particularly on complex reasoning benchmarks.

Opening Kapyn…