Google releases DiffusionGemma: 26B open-weight model generating text 4× faster via parallel diffusion
Google DeepMind released DiffusionGemma on June 10 under Apache 2.0 — a 26B-parameter model that generates full paragraphs simultaneously, image-diffusion-style, rather than token-by-token. It reaches 700+ tokens/s on an RTX 5090 and 1,000+ on a single H100, with support already merged into vLLM, MLX, and Hugging Face Transformers.
why it mattersFirst open-weight production diffusion text model: trades some benchmark accuracy for a 4× throughput multiplier, with day-one inference-stack support.