Google Releases DiffusionGemma, a 26B MoE Text-Diffusion Model With Up to 4x Faster Generation

Google / Hacker News·high signal

Google open-sourced DiffusionGemma on June 10, a 26B-parameter Mixture-of-Experts model (3.8B active) that generates text via diffusion — refining 256 tokens in parallel per forward pass with bidirectional attention instead of left-to-right decoding. It hits 1,000+ tokens/sec on an H100 and 700+ tokens/sec on an RTX 5090, fits in 18GB VRAM quantized, and ships Apache 2.0 on Hugging Face, though Google notes output quality trails standard Gemma 4. NVIDIA published companion RTX acceleration, signaling a real bet on fast local/interactive inference for in-line editing and rapid iteration.