Google DiffusionGemma boosts text speed 4x

Google DeepMind released DiffusionGemma, a 26 billion-parameter open-weights model that uses diffusion techniques from image generators to produce text. The model generates entire paragraphs of tokens simultaneously rather than one token at a time. It runs on consumer hardware with just 18 GB of DRAM or VRAM and delivers up to 4x faster output performance.