Google DeepMind releases DiffusionGemma, a diffusion LLM 4x faster than Gemma 4

Google DeepMind has unveiled DiffusionGemma, a new class of language model that generates text through diffusion rather than the standard autoregressive approach. According to the team, the model runs 4x faster than any other model in the Gemma 4 family at comparable quality.

How a diffusion LLM works

Classic language models such as GPT, Llama and Gemma generate text sequentially: token by token, left to right. Producing 500 tokens takes 500 sequential inference steps. This is a fundamental limit of the architecture: each next token depends on all the previous ones, so parallelism is impossible by the very nature of autoregression.

The diffusion approach works differently. The same idea behind Stable Diffusion and DALL-E for images is now applied to text: the model learns to reconstruct the original text from random noise, refining the entire sequence at once. It does not go left to right but iteratively, in parallel across all positions simultaneously.

• Autoregression: 500 tokens require 500 sequential steps
• Diffusion: 500 tokens are processed in 10–50 steps, regardless of length
• The speedup is nonlinear: the longer the text, the more noticeable the advantage

Many teams have tried to make diffusion work for text generation since 2021. The main problem was quality: diffusion text models long lagged behind autoregressive ones, and their output lost coherence and precision of wording. Judging by Hassabis's statement, DiffusionGemma has overcome that barrier.

Demis Hassabis announced it personally

The CEO of Google DeepMind announced the result himself, which is unusual. Executives at that level typically promote entire products or strategic directions, but rarely single out a specific architectural decision as an occasion for congratulations. Hassabis personally congratulated researcher Brian O'Donoghue and the whole team, calling the work "lightning-fast".

"Great innovation in text diffusion.

DiffusionGemma is lightning-fast, 4x faster than other Gemma 4 models. Can't wait to see what people build with it!" — Demis Hassabis

Important context: the comparison is not against outdated benchmarks but against the current Gemma 4 family, which is itself considered one of the most efficient among open models in its class. A fourfold gain over such a baseline is a substantial architectural result.

The economics of inference are shifting

Generation speed determines both API cost and the latency of the end product. If DiffusionGemma generates 4x faster at comparable quality, that opens up several practical possibilities:

• Lower inference cost: less GPU time per response
• Long contexts without an exponential rise in delays
• Competitiveness in latency-sensitive scenarios: chatbots, autocomplete, agentic pipelines
• Potential for unification with diffusion-based image and audio generation

The multimodal synergy is especially interesting: combining text diffusion with the already mature approaches for images and sound yields a single architecture that handles every modality with one principle. Google is already moving in this direction with the Gemini series, and DiffusionGemma looks like a first step toward full diffusion-based multimodality.

What this means

Diffusion LLMs are no longer an academic experiment. When the CEO of one of the world's largest AI labs personally announces an architectural breakthrough, the market reacts. If DiffusionGemma's speed figures are confirmed in independent tests, they may change pricing in the LLM inference market and push competitors to accelerate their own diffusion work. Developers who haven't studied this architecture yet should do so now.