Google Releases DiffusionGemma, a Text Model That Generates Four Times Faster

Google DeepMind has released DiffusionGemma, an open language model that breaks with the approach that has dominated text generation for years. Instead of writing a response token by token, the model produces an entire block of text at once and refines it over subsequent passes, letting it generate responses up to four times faster than standard autoregressive models.

The model is built on the Gemma 4 architecture in its Mixture of Experts variant, but instead of the classic autoregressive mechanism it uses discrete diffusion. The generation process starts with a block of 256 random placeholder tokens resembling noise. In subsequent passes, the model evaluates every position in the block at once, locks in the ones it is most confident about, and re-samples the uncertain spots for another round of evaluation.

How text diffusion works

Classic large language models such as GPT or standard Gemma generate text sequentially, word by word, basing each next token on everything that came before it. DiffusionGemma reverses that logic and uses bidirectional attention, so every position in the generated block can take into account all other positions at once, not just the ones that appeared earlier.

Google notes that this approach works especially well for nonlinear tasks, such as filling in code in the middle of an existing file or formatting text in markdown, where the next part of the content depends not only on what came before but also on what comes after.

The price of speed

The company does not hide that the speedup comes at a cost. In its official announcement, Google states outright that the overall response quality of DiffusionGemma is lower than that of standard, autoregressive Gemma 4, and recommends the latter model for applications that require maximum precision.

The overall output quality of DiffusionGemma is lower than standard Gemma 4 - Google DeepMind statement at the model's launch

DiffusionGemma makes the most sense where an instant response from the model matters more than maximum quality in every answer. Google points to use cases such as real-time text editing, quickly iterating on a draft, or working with content that has a nonlinear structure. The company also notes that the speedup works best with local inference and a low number of parallel requests, not with handling heavy cloud traffic at scale.

Hardware and availability

The model has been added to the vLLM ecosystem as the first diffusion model natively supported by that serving platform. Tests in this environment showed performance of around 1,008 tokens per second on an H100 at FP8 precision, and 1,288 tokens per second on the newer H200 card. Google also cautions that Macs with Apple Silicon chips may not achieve a similar speedup, due to the different ratio of compute power to memory bandwidth in the unified memory architecture.

DiffusionGemma is not the first commercial diffusion model for text generation, but it is by far the largest one released publicly using this approach. A similar technique, though on a smaller scale and aimed mainly at coding tasks, was previously showcased by the startup Inception Labs with its Mercury Coder model.

What it means for developers

For app developers, this is a new addition to Google's portfolio of open models alongside Gemma 4 itself, not a replacement for it. The model's authors are Google researchers Brendan O'Donoghue and Sebastian Flennerhag, and releasing the weights under the Apache 2.0 license lets anyone download and run the model locally, including on consumer cards such as the RTX 4090 and 5090, without needing a paid API.

For Polish companies and developers already using Google's open models, such as earlier versions of Gemma, DiffusionGemma is first and foremost an interesting technical experiment to test in use cases where response time matters, for example in interactive text or code editing tools running locally, rather than a ready-made replacement for production deployments that require the highest quality generated text.