DiffusionGemma 26B A4B
Hugging Face (open weights) Open weights
The catch: An experimental block-diffusion model: the card only shows a Transformers example (DiffusionGemmaForBlockDiffusion), though vLLM's supported-models list now includes it; there is no Ollama tag.
| Access | Open weights |
|---|---|
| Free limits | Free to download and self-host; you pay only for your own compute. 25.2B total / 3.8B active (MoE); generates text by discrete diffusion over 256-token blocks for faster decoding; reads text and images; up to 256K-token context. Released 9 Jun 2026. |
| Modality | text, code, image |
| Credit card | Not required |
| Commercial use | Allowed |
| Trains on your data | – |
| Licence or terms | Apache-2.0 |
| Context window | 262,144 tokens |
| Model id | google/diffusiongemma-26B-A4B-it |
| Base URL | – |
| Last checked | 8 Oct 2026 |
How to use DiffusionGemma 26B A4B
Download the weights and serve them yourself; the quickstart below shows one way. Licence: Apache-2.0.
Quickstart
pip install -U transformers torch accelerate # then load google/diffusiongemma-26B-A4B-it with DiffusionGemmaForBlockDiffusion.from_pretrained, as in the model card
Paste into Claude Code
I want to run the open-weight model "DiffusionGemma 26B A4B" (google/diffusiongemma-26B-A4B-it) for my project. Please: 1. Check the model card and tell me the hardware it needs and the simplest way to serve it (vLLM, SGLang, Ollama or a hosted endpoint). 2. Give me a minimal working setup and a test request. 3. Tell me the licence terms for commercial use. 4. If a cheaper or smaller variant would do the job, mention it.
Before you ship on it
- An experimental block-diffusion model: the card only shows a Transformers example (DiffusionGemmaForBlockDiffusion), though vLLM's supported-models list now includes it; there is no Ollama tag.
- Apache-2.0 (licence tag on the HF repo, linking the Gemma 4 licence page): use, modify, distribute and sell, including commercially; keep the licence and notices when you redistribute.
- Free to run means you pay for the GPUs, the serving stack and the on-call.
Questions
Is DiffusionGemma 26B A4B free?
Yes. Open weights offers it as open weights, with these limits: Self-host free · Apache-2.0. No credit card is needed.
Can I use DiffusionGemma 26B A4B commercially?
Yes, under Apache-2.0. Check the current terms before you ship.
How do I start with DiffusionGemma 26B A4B?
Download the weights and serve them yourself; the quickstart below shows one way. Licence: Apache-2.0.
Source: huggingface.co/google/diffusiongemma-26B-A4B-it, checked 8 Oct 2026.