Google's DiffusionGemma Writes Text a Different Way

Almost every chatbot you have used writes the same way: one token at a time, left to right, each word chosen with the ones before it in view. Google DeepMind's latest open release questions that habit. DiffusionGemma, published on 10 June under a permissive Apache 2.0 licence, generates text by a different method borrowed from image models, and the payoff is speed.

The technique is called diffusion. Instead of building a sentence word by word, the model starts with a block of garbled "noise" tokens and refines them in repeated passes until coherent text emerges, working on a span of 256 tokens at once rather than marching through them in order. Image generators have used this trick for years. Applying it cleanly to language has been harder, because text is discrete and unforgiving in a way that pixels are not.

Fast, small, and honest about the trade

The numbers are the draw. Google says DiffusionGemma produces more than 1,000 tokens per second on a single Nvidia H100, up to four times faster than comparable word-by-word models, as MarkTechPost reported. It is a mixture-of-experts model with 26 billion parameters in total but only about 3.8 billion active at a time, which keeps it light enough to run on a single graphics card with roughly 18 GB of memory. Nvidia has already tuned it for local use on consumer hardware.

What earns the company some trust here is that it does not oversell the result. On its own published benchmarks, DiffusionGemma scores below the standard Gemma 4 on quality. As The Register put it, this is not another chatbot so much as a different way of producing text, one that gives up a little accuracy for a large gain in speed. That is a reasonable bargain for plenty of jobs, from drafting and autocomplete to bulk processing where throughput matters more than polish.

Why an experiment is worth watching

DiffusionGemma is best read as a research bet, not a finished product. The autoregressive approach behind today's leading systems is mature and very good, and one experimental model will not displace it. But the appetite for faster, cheaper generation is real, and diffusion is one of the more credible routes to getting there. By releasing the weights openly rather than locking the idea behind an API, DeepMind is inviting the wider community to poke at it.

It also keeps the pressure on locally run models, the same ground Google staked out with Gemma 4 on laptops and that open-weight rivals like Moonshot's new Kimi coder keep contesting. If diffusion text models grow up, the way our software writes could end up looking quite different from the steady left-to-right stream we have grown used to.