--- title: DiffusionGemma: 4x faster text generation | daily.dev description: Google has released DiffusionGemma, an experimental 26B Mixture of Experts open model that uses text diffusion instead of autoregressive token-by-token... canonical: https://daily.dev/posts/diffusiongemma-4x-faster-text-generation-yowvfywbx twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: DiffusionGemma: 4x faster text generation | daily.dev og:description: Google has released DiffusionGemma, an experimental 26B Mixture of Experts open model that uses text diffusion instead of autoregressive token-by-token... og:url: https://daily.dev/posts/diffusiongemma-4x-faster-text-generation-yowvfywbx og:image: https://api.daily.dev/og/posts/YOwVfYwBx.png og:image:alt: DiffusionGemma: 4x faster text generation og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # DiffusionGemma: 4x faster text generation **[DeepMind](https://daily.dev/sources/dm)** · 5 min read · 1 upvotes · 0 comments ## Summary Google has released DiffusionGemma, an experimental 26B Mixture of Experts open model that uses text diffusion instead of autoregressive token-by-token generation. By generating 256 tokens simultaneously in parallel, it achieves up to 4x faster inference on GPUs — over 1000 tokens/second on an H100 and 700+ on an RTX 5090. The model activates only 3.8B parameters during inference, fitting within 18GB VRAM when quantized. Key advantages include bi-directional attention (useful for code infilling, inline editing, and non-linear text tasks) and iterative self-correction. Trade-offs include lower output quality compared to standard Gemma 4 and diminishing returns in high-concurrency cloud serving. The model is released under Apache 2.0 and supports vLLM, MLX, Hugging Face Transformers, Unsloth, and NVIDIA NeMo, with llama.cpp support coming soon. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is DiffusionGemma and how is it different from standard Gemma 4? DiffusionGemma is an experimental 26B Mixture-of-Experts model from Google that generates text using diffusion instead of token-by-token autoregressive decoding, producing 256 tokens per forward pass with bi-directional attention. It activates only 3.8B parameters during inference, fits within 18GB VRAM when quantized, and delivers up to 4x faster generation on GPUs, though with lower output quality than standard Gemma 4. _Developers weighing speed against output quality for local inference can track model releases like this via daily.dev._ ### How fast is DiffusionGemma on an NVIDIA H100 or RTX 5090 compared to typical autoregressive LLMs? DiffusionGemma generates over 1000 tokens per second on a single NVIDIA H100 and over 700 tokens per second on an NVIDIA GeForce RTX 5090, roughly 4x faster than typical autoregressive decoding on the same hardware. The speedup is strongest for local, low-concurrency, single-accelerator use rather than high-QPS cloud serving, where autoregressive batching remains more cost-effective. _Anyone benchmarking local inference throughput can follow model performance comparisons like this through daily.dev._ ### What tools support running or fine-tuning DiffusionGemma? DiffusionGemma can be served with MLX, vLLM (with Red Hat integration), and Hugging Face Transformers, and fine-tuned using Unsloth, NVIDIA NeMo, or Google's Hackable Diffusion, a modular JAX toolbox built for composability. Official llama.cpp support is coming soon, and NVIDIA optimized the model across RTX 5090/4090, Hopper, Blackwell, DGX Spark, DGX Station, and RTX PRO hardware. _Developers picking a serving stack for new open models can keep track of integration support through daily.dev._ ## Similar posts on daily.dev - [Google’s DiffusionGemma is 4x faster than its other Gemma models](https://daily.dev/posts/google-s-diffusiongemma-is-4x-faster-than-its-other-gemma-models-vjw4zukla) · The New Stack · 2 upvotes · 0 comments - [Google unveils DiffusionGemma, an AI model that breaks free of left-to-right processing](https://daily.dev/posts/google-unveils-diffusiongemma-an-ai-model-that-breaks-free-of-left-to-right-processing-03hzpmlaz) · InfoWorld · 0 upvotes · 0 comments - [NVIDIA Accelerates Google DeepMind’s DiffusionGemma for Local AI](https://daily.dev/posts/nvidia-accelerates-google-deepmind-s-diffusiongemma-for-local-ai-xcxwdxism) · NVIDIA · 0 upvotes · 0 comments - [I tried Google's new DiffusionGemma, and watching it generate text like an image is unlike any local LLM](https://daily.dev/posts/i-tried-google-s-new-diffusiongemma-and-watching-it-generate-text-like-an-image-is-unlike-any-local-0op3vd8nm) · XDA Developers · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-inference](https://daily.dev/tags/ai-inference), [#gemma](https://daily.dev/tags/gemma), [#diffusion-models](https://daily.dev/tags/diffusion-models), [#text-generation](https://daily.dev/tags/text-generation) [View this post on daily.dev](https://daily.dev/posts/diffusiongemma-4x-faster-text-generation-yowvfywbx) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"DiffusionGemma: 4x faster text generation","url":"https://daily.dev/posts/diffusiongemma-4x-faster-text-generation-yowvfywbx","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/diffusiongemma-4x-faster-text-generation-yowvfywbx"},"datePublished":"2026-07-21T16:31:38.617Z","dateModified":"2026-09-13T21:54:59.370Z","description":"Google has released DiffusionGemma, an experimental 26B Mixture of Experts open model that uses text diffusion instead of autoregressive token-by-token...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/cd3dd98a57c2421a1f689ee9037f340d?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/cd3dd98a57c2421a1f689ee9037f340d?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"DeepMind","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"DeepMind","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/dm","url":"https://daily.dev/sources/dm"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/diffusiongemma-4x-faster-text-generation-yowvfywbx","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-inference,gemma,diffusion-models,text-generation","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"DeepMind","item":"https://daily.dev/sources/dm"},{"@type":"ListItem","position":3,"name":"DiffusionGemma: 4x faster text generation"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/diffusiongemma-4x-faster-text-generation-yowvfywbx#faq","mainEntity":[{"@type":"Question","name":"What is DiffusionGemma and how is it different from standard Gemma 4?","acceptedAnswer":{"@type":"Answer","text":"DiffusionGemma is an experimental 26B Mixture-of-Experts model from Google that generates text using diffusion instead of token-by-token autoregressive decoding, producing 256 tokens per forward pass with bi-directional attention. It activates only 3.8B parameters during inference, fits within 18GB VRAM when quantized, and delivers up to 4x faster generation on GPUs, though with lower output quality than standard Gemma 4. Developers weighing speed against output quality for local inference can track model releases like this via daily.dev."}},{"@type":"Question","name":"How fast is DiffusionGemma on an NVIDIA H100 or RTX 5090 compared to typical autoregressive LLMs?","acceptedAnswer":{"@type":"Answer","text":"DiffusionGemma generates over 1000 tokens per second on a single NVIDIA H100 and over 700 tokens per second on an NVIDIA GeForce RTX 5090, roughly 4x faster than typical autoregressive decoding on the same hardware. The speedup is strongest for local, low-concurrency, single-accelerator use rather than high-QPS cloud serving, where autoregressive batching remains more cost-effective. Anyone benchmarking local inference throughput can follow model performance comparisons like this through daily.dev."}},{"@type":"Question","name":"What tools support running or fine-tuning DiffusionGemma?","acceptedAnswer":{"@type":"Answer","text":"DiffusionGemma can be served with MLX, vLLM (with Red Hat integration), and Hugging Face Transformers, and fine-tuned using Unsloth, NVIDIA NeMo, or Google's Hackable Diffusion, a modular JAX toolbox built for composability. Official llama.cpp support is coming soon, and NVIDIA optimized the model across RTX 5090/4090, Hopper, Blackwell, DGX Spark, DGX Station, and RTX PRO hardware. Developers picking a serving stack for new open models can keep track of integration support through daily.dev."}}]} ```