DiffusionGemma borrows the denoising logic of image diffusion for text. Instead of committing one token at a time, it starts with a noisy 256-token canvas and repeatedly locks in confident words while revising the rest.
Google says the model can average 15 to 20 tokens per forward pass, exceed 1,000 tokens per second on one H100 and run above 700 tokens per second on an RTX 5090 after quantization. The 26B MoE activates about 3.8B parameters per step and supports long context.
Google is careful about the trade-off: overall quality is below standard Gemma 4, so the model is aimed at local coding edits, fast iteration, document analysis and constrained generation rather than high-quality cloud serving.
Sources：MarkTechPost、CocoLoop、MLQ News.