Gemini Diffusion

Gemini Diffusion, unveiled at Google I/O on May 20, 2025, is Google DeepMind's research foray into diffusion-based language modeling. Instead of generating tokens left to right, it starts from noise and iteratively denoises whole blocks of text in parallel, which also enables mid-generation self-correction of uncertain positions.

DeepMind reported speeds of 1,479 tokens per second (up to 2,000 on code tasks), with benchmark performance roughly matching Gemini 2.0 Flash-Lite: near-identical HumanEval and MBPP scores, slightly ahead on LiveCodeBench, but behind on GPQA Diamond and multilingual MMLU. It launched as an experimental demo with an early-access waitlist and no arXiv paper; the later open-weights DiffusionGemma brought the research to downloadable form.