Google converts Gemma 4 into a diffusion language model

Google DeepMind retrofitted its Gemma 4 model into a text diffusion model, called DiffusionGemma, using less than 10% of the original training budget. It generates 256 tokens in parallel at about 1,500 tokens per second, though quality still trails the original autoregressive model, especially on reasoning tasks.

Why it matters: Diffusion language models promise much faster generation than token-by-token decoding, and this shows a cheap path to get one by retrofitting an existing model instead of training from scratch. The reasoning-quality gap is the real open question determining whether diffusion LLMs become a mainstream alternative or stay a speed-focused niche.