• DiffusionGemma is Google's new experimental open model built on the Gemma 4 architecture. • Unlike traditional LLMs that generate text token-by-token, it generates and refines blocks of text in parallel using a diffusion-based approach. • Google says it can deliver up to 4× faster inference on dedicated GPUs. • The model activates ~3.8B parameters per step from a 26B-parameter Gemma 4 MoE architecture. • Released under the Apache 2.0 license with support for local deployment and ...