Google DeepMind releases DiffusionGemma, a block-generation open model
Google DeepMind's DiffusionGemma generates text in parallel blocks rather than word-by-word, delivering up to 4x faster output on dedicated GPUs.
Google DeepMind has released DiffusionGemma, an experimental open model that departs from the standard autoregressive approach used by most LLMs. Rather than predicting one token at a time, DiffusionGemma generates entire blocks of text simultaneously — a diffusion-based method that allows the model to self-correct within a generation pass and format complex markdown in real time.
The practical headline is speed: Google DeepMind claims up to 4x faster output on dedicated GPUs compared to token-by-token generation. For developers building latency-sensitive applications — chat interfaces, coding assistants, document generation pipelines — block-level generation could reduce inference costs or unlock use cases where streaming word-by-word output is a bottleneck. As an open model, DiffusionGemma is also available for direct experimentation, giving teams a concrete alternative architecture to benchmark against standard transformer-based generation.