IA · 11 June 2026 · 3 min read
Google Releases DiffusionGemma: Text Generation Speeds Up 4x with Parallel Diffusion
In brief: Google DeepMind has launched DiffusionGemma, an open-source model that changes how AI generates text. Instead of relying on the traditional autoregressive method (predicting text word-by-word), it utilizes a diffusion process similar to image generators. This allows DiffusionGemma to generate complete text blocks in parallel, offering up to 4x faster performance on local hardware such as GPUs.
by Team Mocchi's
Beyond the Autoregressive Paradigm: How Text Diffusion Works
Until now, the vast majority of Large Language Models (LLMs) have relied on an "autoregressive" approach. This means the AI generates text linearly, from left to right, predicting and writing one word (or "token") at a time. While highly effective, this sequential generation creates major bottlenecks tied to memory bandwidth, as hardware must reload the model's parameters for every single token produced.
As reported by Ars Technica, Google DeepMind has released a new open-source model that breaks away from this framework: DiffusionGemma. Instead of writing sequentially, DiffusionGemma employs a technique similar to image generators like Midjourney or Stable Diffusion. The model starts by populating a digital canvas with a field of "placeholder tokens" and iteratively "denoises" the text, producing a complete block of text in parallel.
Technical Specifications and Local Hardware Performance
DiffusionGemma is a new addition to the Gemma 4 open model family and is built as a Mixture of Experts (MoE) model. While it has a total of 26 billion parameters, only 3.8 billion are activated during inference. This optimized design allows the model to run comfortably on local hardware with at least 18 GB of VRAM, making it accessible for developers and enterprise workstations.
Initial benchmarks show impressive processing speeds:
- Running on a high-end consumer GPU like an Nvidia RTX 5090, the model achieves a throughput of approximately 700 tokens per second.
- Utilizing an enterprise-grade accelerator like the Nvidia H100, DiffusionGemma surpasses 1,000 tokens per second.
This represents a speedup of roughly four times compared to similarly sized autoregressive Gemma models. By shifting the hardware bottleneck from memory bandwidth to raw compute, the model can process and generate up to 256 tokens in parallel in a single pass.
Practical Advantages: From Editing to Non-Linear Problem Solving
Parallel text generation unlocks use cases where sequential models have historically struggled. In traditional LLMs, each word depends strictly on the tokens that came before it. This linear constraint makes it difficult to solve non-linear problems or tasks requiring continuous global evaluation of context.
A practical showcase from Google DeepMind involves solving Sudoku puzzles. This task is notoriously difficult for standard LLMs because the placement of any number depends on both past and future constraints on the grid. DiffusionGemma, through its ability to continuously evaluate and self-correct large blocks of tokens across the entire canvas, handles these puzzles with ease. Other ideal applications include in-line text editing (where the AI modifies sections of a document while maintaining total coherence with the surrounding text), molecular sequencing, and mathematical graphing.
Current Limitations: Why Autoregressive Models Dominate the Cloud
Despite this performance leap, text-based diffusion has specific limitations, explaining why Google has not yet migrated its flagship cloud models like Gemini to this architecture.
The primary hurdle is the error rate. In image generation, a slightly misaligned pixel is often unnoticeable. In language, however, data is discrete: a single incorrect token can make an entire sentence meaningless, requiring the system to discard and regenerate the block. Additionally, for very short outputs (such as a simple five-token response), diffusion models are less efficient because they must perform multiple denoising passes, whereas an autoregressive model can complete the task in just five quick sequential steps.
Nonetheless, the release of DiffusionGemma marks an important milestone for local AI development. It proves that scaling efficiency is not just about increasing parameter counts, but about exploring entirely new computational architectures.