
Google Releases DiffusionGemma, an Open Model That Generates Text From Noise 4x Faster
Google DeepMind dropped DiffusionGemma, a 26B mixture-of-experts open-weights model (3.8B active) that generates 256-token blocks in parallel via text diffusion instead of one word at a time. It hits 1,000+ tokens/sec on an H100 and fits in 18GB of VRAM, with day-zero support in vLLM, Transformers, and MLX under Apache 2.0. It matters because it's the first serious open-weight bet that diffusion, not autoregression, could be the faster path for interactive local AI, with bidirectional attention enabling real-time self-correction.







