DiffusionGemma: Google's 1,500-token-per-second diffusion language model

DiffusionGemma Technical Report

Google's DiffusionGemma generates around 1,500 output tokens per second on a single H100 GPU by refining 256-token blocks in parallel, avoiding autoregressive decoding. The open-weight model fine-tunes Gemma 4's mixture-of-experts architecture with 3.8B activated parameters, using under 10% of the original training token budget. It establishes a new Pareto frontier for speed versus capability, retaining thinking mode, multimodal inputs, and long contexts.