Intel / article

DiffusionGemma: 4x faster text generation

Source: deepmind.google
Date: Published Jun 10, 2026

Context

Google DeepMind's experimental DiffusionGemma generates text by repeatedly refining blocks of 256 , allowing positions throughout a block to inform one another. Its announcement reports up to fourfold faster generation on dedicated GPUs. Processing a larger block together gives the hardware more work at once, particularly when serving a small number of requests.

That makes the design relevant to responsive local editing and code infilling, the uses Google names. Google also reports lower overall output quality than standard Gemma 4 and diminishing speed advantages when conventional models already keep cloud hardware busy with many simultaneous requests. The useful comparison therefore depends on workload and required quality.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Read the source deepmind.google