# Z-Image-Turbo — card for AI agents

Efficient open-weights image generation model built on a single-stream diffusion transformer.

Category: AI image generators · maker Tongyi-MAI · checked 2026-10-27

## Price
- Price: Free
- Free tier: Open weights, free to download; self-hosted.

## License
Apache-2.0 — https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/blob/main/README.md

- commercial: **yes** (Commercial use OK)
  > license: apache-2.0
  (https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/blob/main/README.md)
- in app: **yes** (Can ship inside your app)
  > license: apache-2.0
  (https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/blob/main/README.md)

## Rules to follow
- Keep the Apache-2.0 license and notices when you redistribute the weights.
- The model card does not add a separate output license; check the repo before shipping.

## How to connect
- Download only
- Repository: https://github.com/Tongyi-MAI/Z-Image
- Website: https://huggingface.co/Tongyi-MAI/Z-Image-Turbo

## Notes
Model card also lists an online demo Space and ModelScope mirror.

## Numbers (AgentGid data)
- Gid Score: 35
- GitHub stars: 12,167

## Lines to paste into your prompt or AGENTS.md

```
- Z-Image-Turbo (AI image generators). Efficient open-weights image generation model built on a single-stream diffusion transformer. Price: free; commercial use OK.
  Connect: download from https://huggingface.co/Tongyi-MAI/Z-Image-Turbo (no API).
  Rule: Keep the Apache-2.0 license and notices when you redistribute the weights.
  Rule: The model card does not add a separate output license; check the repo before shipping.
  License: https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/blob/main/README.md (checked 2026-10-27; details: https://agentgid.com/assets/z-image-turbo/)
```

## Links
- AgentGid page: https://agentgid.com/assets/z-image-turbo/
- Is it free for commercial use? https://agentgid.com/assets/z-image-turbo/commercial-use/
- Source: https://huggingface.co/Tongyi-MAI/Z-Image-Turbo
- Source: https://github.com/Tongyi-MAI/Z-Image


## Z-Image-Turbo: fast open model for realistic pictures

Z-Image-Turbo is a free image model from Alibaba's Tongyi team that makes pictures from a text description. It is tuned for realistic photos and can write English and Chinese text inside the picture. It runs on your own computer with a gaming-class graphics card, and the Apache-2.0 license allows paid use.

### Where to use it
- Realistic photos for a site or app: Make photo-like pictures from a description, such as a cozy cafe interior for a landing page, on your own machine without paying per image.
- Posters and banners with lettering: Ask for a picture that includes a short headline or sign text in English or Chinese, for example a poster or a shop sign.
- Images on a gaming graphics card: Run it at home or in the office on a graphics card with 16 GB of memory; with memory saving turned on, smaller cards can work too, only slower.
- Fast image step in your own service: An app that makes images for its users calls the model on its own server; only 8 model steps per picture keep each request short.
- Blog and article cover images: Make a matching cover picture for each new post from its title and a short style line, then pick the best of a few seeds. (idea, not from the official pages)
- Product mockups for a pitch: Show a product idea in a realistic setting, such as a bottle on a kitchen counter, before any real photo shoot. (idea, not from the official pages)

### How to start
- Install a Diffusers version that includes Z-Image support (the README installs it from source).

```
pip install git+https://github.com/huggingface/diffusers
```

- Load the Turbo model in bfloat16 and move it to the graphics card.

```
pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=False,
)
pipe.to("cuda")
```

- Generate with 9 steps and guidance 0, then save the picture.

```
image.save("example.png")
```


### License
Apache-2.0 license, so paid use is allowed. If you pass the model files on, include the license and keep the notices; the model card sets no separate rules for the pictures you make.

### Good to know
- The download is large, about 33 GB of model files, so plan disk space and the first download time.
- The Turbo version trades variety for speed and is not meant for further training. The base Z-Image model is the one for training your own style.
- The official examples expect an NVIDIA graphics card; memory saving helps on smaller cards.
- Speed and ranking numbers on the pages come from the makers and leaderboards they cite, not from independent tests.

Checked 2026-11-10 against: https://huggingface.co/Tongyi-MAI/Z-Image-Turbo, https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/raw/main/README.md, https://huggingface.co/api/models/Tongyi-MAI/Z-Image-Turbo, https://huggingface.co/api/models/Tongyi-MAI/Z-Image-Turbo/tree/main?recursive=1, https://github.com/Tongyi-MAI/Z-Image/blob/main/README.md, https://github.com/Tongyi-MAI/Z-Image/blob/main/LICENSE, https://github.com/Tongyi-MAI/Z-Image/commits/main.atom

### Mistakes to avoid
- Don't: Using guidance_scale 3.5 to 7 and 30 to 50 steps as with other image models → Do: num_inference_steps=9 and guidance_scale=0.0 for Turbo (The Turbo model is distilled for 8 steps without guidance.)
- Don't: Loading with StableDiffusionPipeline or an old diffusers release → Do: Use ZImagePipeline from a diffusers version that includes Z-Image support (Z-Image support was added to diffusers recently, so older releases do not have ZImagePipeline.)
- Don't: Calling pipe.to("cuda") on a card with less than 16 GB and getting out-of-memory errors → Do: Call pipe.enable_model_cpu_offload() instead of pipe.to("cuda") (Offloading keeps only the active part of the model on the card.)
- Don't: Using Turbo as the base for LoRA or style training → Do: Train on the base Z-Image model, which the makers mark as easy to fine-tune (The model table lists fine-tunability for Turbo as N/A.)

### Questions
- Can I use the pictures in a paid product? The model is under Apache-2.0, which allows commercial use, and the model card adds no separate rules for outputs.
- Can I try it before installing? Yes, the makers link an online demo on Hugging Face Spaces and on ModelScope.
- Why is guidance set to 0? The Turbo model is distilled to work without guidance; the official example says guidance should be 0 for it.
- Can it edit an existing picture? Not this model. Editing is the job of a separate Z-Image-Edit model, which the model table lists as not yet released.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
