DiffusionGemma: 4x faster text generation
Summary
Google introduces DiffusionGemma, an experimental open model that uses diffusion for text generation to reach up to 4x faster inference on GPUs. The 26B Mixture of Experts model generates text in parallel blocks, targets speed-critical local workflows, and is released under an Apache 2.0 license, with trade-offs in output quality compared to Gemma 4. The article covers hardware optimizations, fine-tuning possibilities, and practical guidance for developers.