A fine-tune of Gemma 4, DiffusionGemma abandons one-token-at-a-time decoding for discrete diffusion — refining a 256-token block at once and topping 1,000 tokens per second on a single H100.
2026-08-20 · by Moe Ameen
Google DeepMind published a technical report for DiffusionGemma on arXiv on July 31, 2026 (2608.00146) and released the model as open weights. It is an experimental large language model that swaps the usual autoregressive, left-to-right decoding for discrete diffusion: instead of predicting one token at a time, it starts from a masked block of up to 256 tokens and iteratively denoises the whole block in parallel over a small number of refinement steps.
DiffusionGemma is built by fine-tuning Gemma 4, Google's open multimodal model. The base is a mixture-of-experts model with about 26 billion total parameters that activates roughly four billion per step — the report specifies 3.8B activated of 25.2B total, and the public release is named diffusiongemma-26B-A4B-it. Training used two stages: supervised fine-tuning to teach bidirectional denoising, then reinforcement learning paired with sampler distillation to improve quality and reduce the step count. Google says it reused less than 10% of the original model's training-token budget.
The reason to care is throughput. DiffusionGemma generates around 20 tokens per forward pass and exceeds 1,000 tokens per second on a single NVIDIA H100 — up to four times the token output of comparable autoregressive decoding, and faster than autoregressive models even with speculative decoding. Adaptive stopping lets it terminate a block early, well under its 48-step maximum. Because it inherits from Gemma 4, it keeps thinking mode, long context, and multimodal inputs (text, images, and video in; text out), and it can still fall back to ordinary token-by-token generation with only minor quality loss — which the team frames as a path toward hybrid diffusion-autoregressive decoding.
The weights are downloadable from Hugging Face, Kaggle, and Google Cloud's Vertex AI Model Garden, and a quantized build fits within the 24GB of a consumer GPU (an RTX 4090 or 5090), so it runs on a single card. Google positions it as research-grade and tuned for local, low-concurrency use rather than high-throughput cloud serving. As with any experimental release, treat the specifics as an early snapshot and confirm them in the model card.
The honest read on DiffusionGemma is that it makes the cheapest part of content even cheaper. Drafting was already the fast step; now it is nearly free. What it does not touch is the expensive part — turning a draft into finished, on-brand video, images, and long-form and getting it onto every platform on a schedule. That is the work [Kompozy](/) exists to do, and it is why a faster text model changes your input options without changing where the leverage is.
If you enjoy running open models, self-host DiffusionGemma to spin out raw angles and scripts, then drop the best one into Kompozy as a source: it fans that single input into captioned [Persona Shorts](/glossary/persona-shorts) and avatar video, brand-exact [Carousels](/glossary/hyperframes), Quote Graphics, a Blog Article, and an Email Newsletter — all rewritten under one [Persona Brief](/glossary/persona-brief) with a banned-word filter, reviewed per post, then scheduled and published across the eight social platforms plus blog and email. If you would rather not babysit a GPU at all, you do not have to: Kompozy already runs managed Claude and OpenAI for its copy, with a bring-your-own-key option on the Founding tier. Either way, the model is an interchangeable front end — the durable value is the on-brand generation and multi-platform [distribution](/glossary/content-repurposing) that a raw LLM leaves entirely to you. For other fast, open options in the same family, see [Gemini 3.7 Flash](/news/gemini-3-7-flash-launch) and [DeepSeek's V4 pricing shift](/news/deepseek-api-pricing-update).
No. The name is misleading. DiffusionGemma outputs text only — the "diffusion" refers to how it generates that text, by denoising a whole block of tokens in parallel instead of writing one token at a time. It can take images and video as input and answer questions about them, but everything it produces is written language.
Speed and how it decodes. A standard LLM writes left to right, one token at a time. DiffusionGemma refines a block of up to 256 tokens simultaneously, reaching over 1,000 tokens per second on a single H100 — up to four times faster token output than comparable autoregressive decoding. It is built on Gemma 4 and keeps thinking mode, long context, and multimodal input.
Yes. It was released as open weights on Hugging Face, Kaggle, and Vertex AI Model Garden, and a quantized version fits within the 24GB of a consumer GPU (an RTX 4090 or 5090), so it runs on a single card. Google describes it as research-grade and best suited to local, low-concurrency use rather than high-throughput cloud serving.
It makes drafting nearly free, but drafting was never the hard part — turning a draft into finished, on-brand video, images, and long-form and publishing it everywhere is. A tool like Kompozy takes a raw draft and generates captioned shorts, carousels, a blog, and a newsletter in your brand voice, then schedules them across the eight social platforms plus blog and email.