Google took an ordinary language model and taught it to write 256 words at once — for under a tenth of the original training cost

DiffusionGemma is an open-weight experimental model from Google DeepMind that does not generate text one token after another. It refines blocks of 256 tokens in parallel, the way an image diffusion model sharpens a whole picture at once, producing around twenty tokens per forward pass and roughly 1,500 output tokens per second on a single Nvidia H100. The technical report, credited to a 43-author team and posted to arXiv on 31 July, makes a narrower and more useful claim than the speed number. Diffusion language models have until now been treated as a separate lineage that has to be trained from scratch, which is why almost nobody has one. Google did not train from scratch: it took Gemma 4, a mixture-of-experts model with 25.2 billion total and 3.8 billion active parameters, and converted it — first supervised fine-tuning to reconstruct corrupted blocks of text, then a combined reinforcement-learning and sampler-distillation stage — using fewer than ten per cent of the tokens the original model was trained on. The converted model keeps thinking mode, multimodal input and long context, and the report says it beats autoregressive models on generation speed even when those use state-of-the-art speculative decoding. The weights are on Hugging Face under Apache 2.0. The speed advantage narrows once a server is handling roughly 32 concurrent requests, which is the regime most production deployments actually live in.

✓ Verified · 2 sources

Related stories