Model3h ago
Google Releases DiffusionGemma Text Generation Model
A 26-billion-parameter model called DiffusionGemma generates text by refining noise rather than predicting tokens sequentially. It hits 1,000 tokens per second on a single H100 GPU, quadrupling the speed of autoregressive rivals. However, output quality remains lower. Google positions this as an experimental tool for developers to test non-linear generation.