Z-Image.
Z-Image is an open-source image model from Alibaba (Tongyi-MAI). Windpaint doesn't serve it yet, so there's no model id or price. Request early access to start with the live models.
- Status
- Coming soon
- Task
- Text to image
- Input
- Prompt, optional negative prompt
- Output
- Images
- Developer
- Alibaba (Tongyi-MAI)
- Parameters
- 6B
- Licence
- Apache-2.0
- Released
- January 2026
What the model does.
Paraphrased from the official model card.
Z-Image is the full, undistilled foundation model behind Z-Image Turbo. Where Turbo is distilled to run in a few steps, Z-Image keeps the whole training signal and supports classifier-free guidance and negative prompts.
Its card positions it as a starting point for LoRA training, ControlNet-style structural conditioning and other fine-tuning, and says it gives more variation across seeds than Turbo. It's the same 6B single-stream diffusion transformer architecture.
Coming soon to Windpaint.
Z-Image isn't served, so it has no model id, no price and no request example here. When it goes live, this page gets its tiers, prices and request fields like the live models have.
Want this model? Request early access. We're adding models as we add capacity.
Live for images today: Z-Image Turbo (text to image) and FLUX.2 [klein] 4B (text to image). See Image generation for what the API does now.
Is Z-Image licensed for commercial use?
Licence: Apache-2.0.
Yes. Apache-2.0 allows commercial use, including running the model behind a paid API, with no revenue or user caps.
- Apache-2.0 (card metadata, and the repo LICENSE is the unmodified Apache 2.0 text). No extra terms are added. Z-Image repo, LICENSE
This is our reading of the licence, not legal advice. Read the licence text before you rely on it.
Z-Image hardware requirements.
Published by Alibaba (Tongyi-MAI) for running the weights yourself. These are the developer's figures, not Windpaint's.
- Published VRAM
- Not stated
- Applies to
- No VRAM figure on the card. The example loads bf16 weights.
- The card gives no VRAM figure. Its example loads
ZImagePipelinein bfloat16. Z-Image model card - The official README's Community Works section says DiffSynth-Studio (from ModelScope, a community tool) supports Z-Image LoRA training, full training, distillation training and low-VRAM inference. Z-Image repo
Z-Image native resolution.
What the developer's own code and examples use.
- Recommended resolution: 512 × 512 to 2048 × 2048 in total pixel area, at any aspect ratio. The card's example renders at 720 × 1280. Z-Image model card
Recommended settings for Z-Image.
Steps, guidance and defaults from the developer's card and code, for running the weights yourself.
- Guidance scale 3.0 to 5.0 and 28 to 50 inference steps. The example uses 50 steps, guidance 4 and
cfg_normalization=False. Z-Image model card - CFG normalization: False for general stylism, True for realism. Negative prompts are "strongly recommended for better control". Z-Image repo
- The official demo's UI defaults to 30 steps and guidance 4.0, with CFG normalization off. Z-Image demo Space, app.py
Z-Image prompt guide.
From Alibaba (Tongyi-MAI)'s own prompt guide, prompt tooling and examples.
- Use negative prompts. The README's recommended parameters for Z-Image mark them "strongly recommended for better control". Z-Image repo
- The card's example is a long, detailed Chinese prompt covering the subjects, clothing, quoted text on a garment, lighting, palette and depth of field. Z-Image model card
Z-Image limitations.
As the developer states them.
- Per the card's comparison table, it runs 28 to 50 steps against Turbo's 8, and it skipped the RL stage that Turbo went through. Z-Image model card
Questions.
Can I call Z-Image on Windpaint today?
No. Z-Image isn't served yet, so it has no model id, no price and no entry in the catalog endpoint. For images today, z-image-turbo and flux-2-klein-4b are live. There's no date for Z-Image.
Is Z-Image the same as Z-Image Base?
Yes. Tongyi-MAI publishes it as Z-Image, and its card describes it as the non-distilled base model of the family. Z-Image Turbo is the distilled one.
What's the difference between Z-Image base and Z-Image Turbo?
Z-Image is the undistilled base. Per its card, it supports CFG and negative prompts, runs 28 to 50 steps, is rated higher for diversity, and can be fine-tuned. Turbo is distilled to 8 steps and has no CFG or negative prompts.
What resolution does Z-Image base support?
The card recommends 512x512 to 2048x2048 in total pixel area, at any aspect ratio.
Does Z-Image base support negative prompts?
Yes. The developers mark them as strongly recommended.
What licence is Z-Image base under?
Apache-2.0, so commercial use is allowed.
Is Z-Image open source?
Yes. Its weights are licensed Apache-2.0, an OSI-approved open-source licence.
How do I ask for Z-Image?
Request early access at app.windpaint.ai/register. We're letting people in gradually while we add models and capacity.