Google DeepMind's experimental open-weight diffusion language model — it generates text by refining a whole block of tokens in parallel instead of one at a time, hitting over 1,000 tokens per second on a single H100. Technical report published July 31, 2026.
Last verified · 2026-08-20 · by Moe Ameen
DiffusionGemma is an experimental open-weight language model from Google DeepMind, detailed in a technical report published on arXiv on July 31, 2026 (2608.00146). Unlike a normal large language model, which writes one token at a time from left to right, DiffusionGemma uses discrete diffusion: it starts from a masked block of up to 256 tokens and iteratively denoises the whole block in parallel over a handful of refinement steps. The result is a model that writes paragraphs in one sweep rather than word by word.
It is built by fine-tuning Gemma 4, Google's open multimodal model. The base is a mixture-of-experts model with roughly 26 billion total parameters that activates only about four billion per step (the report gives 3.8B activated of 25.2B total, and the Hugging Face release is named diffusiongemma-26B-A4B-it). Training ran in two stages: supervised fine-tuning to teach bidirectional denoising, then reinforcement learning combined with sampler distillation to sharpen quality and cut the number of steps. Google reports it reused less than 10% of the original autoregressive model's training-token budget.
The headline is speed. DiffusionGemma generates around 20 tokens per forward pass and reaches over 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 early — well under its 48-step maximum — once a block's predictions are confident. 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.
It is released as open weights — 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 frames it as research-grade and tuned for local, low-concurrency use rather than high-throughput cloud serving. Treat the specifics as an early snapshot and confirm them in Google's own model card before you depend on a number.
DiffusionGemma is a drafting engine, not a content engine — and that gap is exactly where [Kompozy](/) picks up. What it gives you is raw text at unusual speed: run it locally on one GPU and you can spin out fifty caption variants, a dozen hook lines, or a rough blog outline in the time an autoregressive model takes to write a few. What it does not give you is a finished post. It has no brand voice locked in, no captions burned onto video, no carousel, no thumbnail, no schedule, and no way to publish. The output is a wall of text that still needs to be turned into content and shipped.
Kompozy is the layer that does that. Paste a DiffusionGemma draft — a script, a set of angles, a rough article — into Kompozy as a source, and one input fans into a full multi-format batch: captioned [Persona Shorts](/glossary/persona-shorts) and avatar video, brand-exact [Carousels](/glossary/hyperframes), Photo Posts and Quote Graphics, a Blog Article, and an Email Newsletter — every asset rewritten and governed by one [Persona Brief](/glossary/persona-brief) with a banned-word filter so your voice stays consistent no matter what the raw draft sounded like. Each piece passes a per-post review gate, then [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email. DiffusionGemma writes the raw words fast; Kompozy turns them into on-brand video, images, and long-form, and puts them everywhere. If you like running open models yourself, note the same pattern works with [Gemma 4](/ai-tools/gemma-4) or a [local Gemma 4 26B engine](/ai-tools/gemma-4-26b-local-engine).
DiffusionGemma is an experimental open-weight language model from Google DeepMind, described in a technical report published July 31, 2026. It uses discrete diffusion to generate text in parallel blocks of up to 256 tokens instead of one token at a time, which makes it very fast. It is a fine-tune of the Gemma 4 mixture-of-experts model — about 26B total parameters with roughly 4B active per step.
No. Despite the "diffusion" name — which people associate with image models — DiffusionGemma outputs text only. The diffusion here refers to how it generates that text (denoising a block in parallel), not to pixels. It can accept images and video as input and answer questions about them, but everything it produces is written language.
Google reports over 1,000 output tokens per second on a single NVIDIA H100, generating roughly 20 tokens per forward pass — up to four times the token output of comparable autoregressive decoding, and faster than autoregressive models even when they use speculative decoding. Adaptive stopping lets it terminate a 256-token block early, well under its 48-step maximum.
It is released as open weights, downloadable from Hugging Face, Kaggle, and Vertex AI Model Garden. A quantized build fits within the 24GB of a consumer GPU (an RTX 4090 or 5090), so it runs on a single card. Google frames it as research-grade and tuned for local, low-concurrency use rather than high-throughput cloud serving.
DiffusionGemma drafts raw text fast; Kompozy turns that draft into finished, on-brand content and publishes it. Paste a script or outline into Kompozy and it generates captioned Persona Shorts, carousels, quote graphics, a blog, and a newsletter in your Persona Brief voice, runs each through a review gate, and schedules them across the eight social platforms plus blog and email.