Google has introduced DiffusionGemma, an experimental open-weights language model that generates text through a diffusion-based process instead of producing one token at a time. Google says the model can reach more than 1,000 tokens per second on a single NVIDIA H100 GPU, though it is being positioned as a developer...
Google introduced DiffusionGemma, an experimental open-weights language model designed to generate text in blocks using diffusion rather than the token-by-token approach common in many large language models.
According to Google’s announcement on The Keyword, DiffusionGemma is a 26-billion-parameter mixture-of-experts model released under the Apache 2.0 license. Google describes it as generating text in blocks rather than sequentially, a design intended to improve inference speed on GPUs.
Google AI for Developers’ model card provides more technical detail, describing DiffusionGemma as a Google DeepMind open-weights model with 25.2 billion total parameters and 3.8 billion active parameters. The model uses discrete diffusion, block-autoregressive multi-canvas sampling, 256-token canvases, and support for context lengths of up to 256,000 tokens, according to the model card.
The key distinction is that DiffusionGemma does not produce text strictly one token after another. Instead, it works with larger blocks of text and refines them through a diffusion-style process. The Decoder compared this approach to the way image generation systems turn noise into a finished picture, while noting that the method is being applied here to text.
Google says DiffusionGemma can deliver up to four times faster GPU text generation compared with comparable autoregressive models. The company’s announcement cites more than 1,000 tokens per second on a single NVIDIA H100 GPU.
NVIDIA’s reported performance figure was also cited by The Decoder, which said the model reaches about 1,000 tokens per second on one H100 and is roughly four times faster than comparable autoregressive models. Decrypt similarly reported Google’s 1,000-token-per-second H100 claim and described the model as generating 256-token blocks simultaneously.
Those speed claims are significant because autoregressive models are constrained by their sequential generation process: each new token generally depends on the tokens generated before it. By generating and refining blocks, diffusion-based text models aim to use parallel hardware more efficiently.
Google is not presenting DiffusionGemma as a finished replacement for mainstream language models. The Decoder reported that the model’s output quality is currently lower than that of comparable autoregressive systems, and that Google is positioning DiffusionGemma as an experimental tool for developers.
Decrypt also noted current runtime limitations in its coverage of the release. Based on Google’s own framing and the third-party reports, DiffusionGemma appears to be aimed at researchers and developers interested in experimenting with faster generation methods, rather than users looking for the strongest general-purpose text model.
The release nevertheless gives developers a concrete open-weights system for testing whether diffusion methods can make text generation faster in practical deployments. Its Apache 2.0 licensing, as described by Google, also makes it more accessible for experimentation than models released under more restrictive terms.
DiffusionGemma reflects a broader research interest in alternatives to standard autoregressive generation. Autoregressive models have dominated modern language AI, but their one-token-at-a-time design can become a bottleneck, especially for long outputs or high-throughput applications.
Google’s release suggests one possible direction: trade some current output quality for faster block-based generation, then improve the technique over time. For now, the most grounded conclusion is narrower than the headline speed figure alone might suggest. DiffusionGemma is an open experimental model with notable performance claims on NVIDIA H100 hardware, but Google and outside coverage both indicate that developers should expect trade-offs while the approach matures.
Google introduced DiffusionGemma, an experimental open weights language model designed to generate text in blocks using diffusion rather than the token by token approach common in many large language models.
A diffusion approach to text generation According to Google’s announcement on The Keyword, DiffusionGemma is a 26 billion parameter mixture of experts model released under the Apache 2.0 license.
Google describes it as generating text in blocks rather than sequentially, a design intended to improve inference speed on GPUs.
Continue reading
