Extremely powerful large language models (LLMs) still operate as though they’re typing on a keyboard, processing workloads in a simple left-to-right fashion. But in locally-run, single-user scenarios, this sequential processing can leave graphics processing units (GPUs) and tensor processing units (TPUs) underutilized.
Google is betting that DiffusionGemma can get around this bottleneck. The new experimental open model generates text “exceptionally fast,” creating entire blocks of text simultaneously through diffusion techniques rather than through token-by-token processing. The company says this technique results in 4x faster inference compared to auto-regressive models that rely on sequential processing.
It can also save users money. Technology analyst Carmi Levy noted that existing pay-per-token monetization models “penalize the use of less than optimally efficient AI solutions.”
But DiffusionGemma “could herald a new generation of task-defined, efficient solutions that can enable expanded compute capacity without draining the operations budget,” he said.
Built on Google’s Gemma 4 family and its Gemini Diffusion research, DiffusionGemma is a 26B mixture-of-experts (MoE) model designed to maximize text output generation.
It essentially shifts how models use hardware, giving processors a larger hunk of work each cycle so it can draft full 256-token paragraphs in sequence. This allows the model to generate text up to 4x faster on GPUs, Google claims. It activates only 3.8B parameters during inference, and, when quantized, can fit within 18GB VRAM on high-end consumer GPUs like Nvidia RTX 5090.
“It upgrades your model inference from a single, sequential typewriter to a massive printing press that stamps the entire block of text simultaneously,” Google research scientists Brendan O’Donoghue and Sebastian Flennerhag wrote in a blog post.
AI image generators begin with pure, random ‘visual noise’ and iteratively refine that into a finalized picture (what’s known as ‘diffusion’); DiffusionGemma applies this same process to text. It does not generate tokens in order, but begins with a “canvas of random placeholder tokens” that it processes in multiple passes, identifying the context tokens it feels are most relevant and using those to refine the rest.
The model has the ability to self-correct, using confidence scoring to re-evaluate tokens in the next pass. “The model iteratively refines its own output, allowing it to evaluate the entire text block at once to fix mistakes in real-time,” O’Donoghue and Flennerhag explained.
DiffusionGemma also has bidirectional attention, they wrote. “Generating 256 tokens in parallel with each forward pass allows every token to attend to all others.” This can be particularly helpful in domains that are non-linear in nature, such as mathematical graphs, code infilling, and in-line editing, they said.
DiffusionGemma is optimized across Nvidia’s hardware stack, making it compatible with consumer setups as well as with high-performance enterprise systems like Hopper and Blackwell.
Because it is released under the Apache 2.0 license, developers can freely use, modify, distribute, and commercialize the software using their preferred tools. It can be run on GPUs or in the cloud through Google Cloud Model Garden or Nvidia NIM, and is available on Hugging Face, GitHub, and vLLM, with support for the open-source library llama.cpp coming soon.
The model is particularly useful in local workflows that are “speed critical,” such as generation of non-linear text structures, and unlocks what Google calls “new patterns of model behavior” like multimodal understanding and generating and rendering code in near real-time.
Levy explained, “DiffusionGemma is particularly well suited for interactive coding and editing where its efficiency allows rapid processing and iterations,” noting that its ability to fit within 18GB of VRAM and its deployability on commonly available local GPUs can potentially benefit customer service-related workloads that lean heavily on real-time interaction and local processing.
“DiffusionGemma also incorporates a thinking mode that is especially adept at problem solving,” he said. For instance, the model was fine-tuned to play Sudoku, a typically challenging task for autoregressive models because each token depends on future tokens. This “rather handily” illustrates the model’s capability to solve more complex problems, Levy noted.
Google freely admits that DiffusionGemma is geared to specific workflows, and there are “key trade-offs.”
The model is engineered for small batch size inferencing and low-latency, high-speed generation low-to-medium batch sizes on a “single capable accelerator.”
In high-QPS cloud serving environments, (where infrastructure is designed to handle tens or hundreds of thousands of requests per second with ultra-low latency), DiffusionGemma’s parallel coding “offers diminishing returns,” and can even result in higher serving costs, Google conceded. In addition, its overall output quality is lower than that of standard Gemma 4, which is built for apps demanding maximum quality.
However, Levy noted that while DiffusionGemma “can be less precise than other models in certain workloads,” subsequent refinement cycles could overcome this limitation.
While Google isn’t sharing runtime costs, it’s clear that this is an efficiency play, he added. “When deployed across the kinds of workloads that would optimally benefit from its architecture, DiffusionGemma seems to have the potential to reduce processing overhead and related costs,” he said.
Taryn Plumb is a freelance writer specializing in AI and cybersecurity. She has also written about data infrastructure, quantum computing, networking hardware and software, and the metaverse. In a previous life she was a news and features reporter for The Boston Globe and numerous other outlets and business journals. She is also the author of several regional history books.
