DiffusionGemma

DiffusionGemma is an open-weights generative text model from Google DeepMind that uses discrete diffusion to generate text in parallel blocks rather than token-by-token, enabling significantly faster output. Built on the Gemma 4 architecture as a 26B MoE model, it accepts multimodal inputs (text, image, video) and is released under Apache 2.0. Source: supabase.