DiffusionGemma
Google's experimental 26B MoE text-diffusion model: 256-token blocks in parallel for up to 4x faster output, but lower quality than Gemma 4.
In short. DiffusionGemma writes text in a different way from normal chatbots. Instead of one word after another, it drafts a whole block of 256 tokens at once and refines it over several passes, which makes answers much faster on a single GPU. Google calls it experimental and says its answers are not as good as those of the standard Gemma 4.
Summary
Google DeepMind released DiffusionGemma on 10 June 2026 as an experimental open model that explores text diffusion. It is built on the 26B A4B mixture-of-experts version of Gemma 4, with 25.2 billion parameters in total, 3.8 billion active and 8 of 128 experts used per token. It reads text, images and video, but not audio, and writes text. The context window is up to 256K tokens, and it has a thinking mode. An autoregressive encoder processes the prompt and caches it, and a decoder with bidirectional attention then denoises a block of 256 tokens at a time, producing 15 to 20 tokens per forward pass.
Google says this gives up to 4 times faster generation on dedicated GPUs, more than 1,000 tokens per second on one H100 and more than 700 on a GeForce RTX 5090, and that it suits in-line editing, rapid iteration, code infilling and non-linear text. The speed claims are for low batch sizes on a single accelerator. Google also says the model's overall quality is lower than standard Gemma 4 and recommends Gemma 4 when quality matters most. The card's own table shows it: 77.6% against 82.6% on MMLU Pro, 69.1% against 88.3% on AIME 2026, and 73.2% against 82.3% on GPQA Diamond. It scores 11.0% against 8.7% on HLE without tools.
The license is Apache 2.0. The BF16 checkpoint is about 52 GB, and Google says a quantized build fits consumer GPUs with 18 to 24 GB of memory. It runs in Hugging Face Transformers, vLLM and MLX, and Unsloth and NVIDIA NeMo are named for fine-tuning. When it launched, llama.cpp support was still described as coming. All benchmark numbers are Google's own.
Good for
- Speed-critical interactive tools such as in-line editing, rapid iteration and code infilling on one GPU
- Researchers and developers exploring text diffusion with an Apache 2.0 model
- Local runs on a high-end consumer GPU with a quantized build, including image and video input
Watch out for
- Google calls it experimental and says its overall quality is lower than Gemma 4, for example 69.1% against 88.3% on AIME 2026
- The speed claims are for small batch sizes on a single GPU, and Google gives no figures for large-batch serving
- It needs recent Transformers or vLLM, and llama.cpp support was announced as coming at launch
- There is no audio input, and all benchmark results are Google's own
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the 26B A4B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| 26B A4B BF16 | 25.2B 3.8B active | 58 GB | 29 GB | 17 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is DiffusionGemma?
- DiffusionGemma writes text in a different way from normal chatbots. Instead of one word after another, it drafts a whole block of 256 tokens at once and refines it over several passes, which makes answers much faster on a single GPU. Google calls it experimental and says its answers are not as good as those of the standard Gemma 4.
- Can I use DiffusionGemma commercially?
- Yes. DiffusionGemma is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is DiffusionGemma open source?
- DiffusionGemma is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run DiffusionGemma on my own computer?
- Yes, on a good graphics card. The smallest variant, 26B A4B, needs about 17 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was DiffusionGemma released?
- DiffusionGemma was released by Google DeepMind on June 10, 2026.
- How long a context window does DiffusionGemma have?
- DiffusionGemma supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Google blog: Introducing DiffusionGemma (10 Jun 2026) blog.google
- DiffusionGemma model card (Google AI for Developers) ai.google.dev
- DiffusionGemma 26B A4B model card (Hugging Face) huggingface.co
- NVIDIA blog: DiffusionGemma for local AI (10 Jun 2026) blogs.nvidia.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Qwen3.6 Alibaba Qwen | 27.8B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Nemotron 3.5 Lightning NVIDIA | 31.6B 3B active | 1M | OpenMDW-1.1 (Permissive) | 22 GB Gaming GPU | |
| LFM2.5-24B-A2B Liquid AI | 24B 2.45B active | 128K | LFM Open 1.0 (Conditional) | 17 GB Gaming GPU |