Z-Image Turbo: specs, license, VRAM, and where your prompt goes
The Morse teamPublished 7 min read
As of October 2026, Z-Image Turbo is a free, open-weight text-to-image model from Alibaba's Tongyi Lab. It's a 6-billion-parameter image transformer that draws a picture in 8 steps, writes legible English and Chinese text inside images, and, by Tongyi's account, "fits comfortably within 16G VRAM consumer devices." It came out on November 26, 2025 under the Apache 2.0 license.
We read Tongyi's Hugging Face model card, its GitHub repository, the technical report on arXiv (version 5, revised July 6, 2026), the project page, and the code of Tongyi's official demo on October 23, 2026. Benchmark figures are Tongyi's, not ours.
Z-Image Turbo specs at a glance
| Spec | Z-Image Turbo | Where Tongyi says it |
|---|---|---|
| Released | November 26, 2025; the checkpoint went up on Hugging Face late on November 25, UTC | GitHub, Hugging Face commits |
| Parameters | 6.15B in the image transformer (30 layers), plus a Qwen3-4B text encoder and the Flux VAE | Technical report, Table 2 and section 4.1 |
| Steps | 8 function evaluations; num_inference_steps=9 "actually results in 8 DiT forwards" | Model card |
| Guidance | guidance_scale=0.0; "Guidance should be 0 for the Turbo models" | Model card |
| VRAM | "Fits comfortably within 16G VRAM consumer devices" | Model card |
| Speed | Sub-second on an Nvidia H800, which needs FlashAttention-3 and torch.compile | Technical report, footnote 1 |
| Text in images | English and Chinese | Model card |
| License | Apache 2.0 | Model card, LICENSE |
| Runs in | Diffusers (ZImagePipeline), Tongyi's own PyTorch code, and ComfyUI | GitHub, Z-Image collection |
| Download | 32.9 GB in Tongyi's repository, by our sum | File list |
Some top search results get these wrong. One dates the release to December 2025, and one help page calls Turbo a 1-step model.
What does "Turbo" mean in Z-Image Turbo?
It means distilled, trained from the base Z-Image model to finish in 8 steps instead of 28 to 50, at the cost of variety. Tongyi's Z-Image card compares them:
- Z-Image (base): 28 to 50 steps, classifier-free guidance, negative prompts, and "high" diversity. Released January 27, 2026.
- Z-Image Turbo: 8 steps, no guidance, no negative prompts, and "low" diversity, but the only variant Tongyi rates "very high" for visual quality. It's also the only one that went through reinforcement learning.
The speed comes from Tongyi's Decoupled-DMD distillation and DMDR, which adds reinforcement learning. Two more variants, Z-Image-Omni-Base and Z-Image-Edit, are still marked "To be released" on the model card as of October 23, 2026, though Tongyi's project page says it is "publicly releasing" both Turbo and Edit.
Is Z-Image Turbo free for commercial use?
Yes. The model card is tagged Apache 2.0, and the GitHub repository carries the full Apache License 2.0, which lets you use, change, and sell what you build, as long as copies you share keep the license and notices and mark your changes. Tongyi adds no terms of its own on top.
The Hugging Face repository holds no LICENSE file, only a license tag, so take the text from GitHub if you need it for a review. This isn't legal advice.
How much VRAM does Z-Image Turbo need?
Tongyi says under 16 GB, but the "6B" in the headlines counts only the image transformer. The pipeline also loads a Qwen3-4B text encoder, which reads your prompt, and a VAE, which turns the result into pixels.
By our sum, Tongyi's repository is 32.9 GB, since it stores the transformer at 32-bit precision. The ComfyUI repackage in Tongyi's own collection has a 16-bit transformer of 12.3 GB and a text encoder of 8.0 GB (Comfy-Org), more than 16 GB together. The card's sample code offers CPU offloading "for memory-constrained devices," Comfy-Org posts 8-bit and 4-bit files, and Tongyi's README says the community engine stable-diffusion.cpp runs Z-Image "on machines with as little as 4GB of VRAM."
How good is Z-Image Turbo at text in images?
Close to the best closed models, by Tongyi's numbers, and far ahead of GPT Image 1 in Chinese. In the technical report:
- CVTG-2K, English text in complex scenes: 0.8585 on average, against 0.8569 for OpenAI's GPT Image 1 at high quality.
- LongText-Bench, longer passages: 0.917 in English and 0.926 in Chinese. GPT Image 1 scored 0.956 in English but 0.619 in Chinese.
- Artificial Analysis leaderboard: an Elo of 1,161, 8th overall and first among open models, which Tongyi's README dates to December 8, 2025.
The report also names limits: "due to limited model size (6B parameters), Z-Image exhibits limitations in world knowledge, intent understanding, and complex reasoning." Its fix, a "Prompt Enhancer" that rewrites prompts with a larger model, isn't in the weights you download.
What does Tongyi say about safety?
The model card has no section on safety, misuse, or limitations. The technical report says Tongyi scored training images for "Not-Safe-for-Work (NSFW)" content and filtered on those scores, which shapes what it learned but blocks nothing at run time.
The run-time filter lives in Tongyi's demo. Its code asks the model's own text encoder whether each prompt is unsafe, using instructions kept in a hidden setting, then runs every finished image through the Stable Diffusion safety checker. If either check trips, you get a placeholder image instead. None of that ships with the weights.
Where can you use Z-Image Turbo, and who sees your prompt?
In three kinds of places: Tongyi's free demo, your own computer, and apps that offer it, such as Morse. The model is the same in each, but who receives your prompt and picture changes.
| Where you run it | Who gets your prompt and image | What their own pages say |
|---|---|---|
| Tongyi's demo on Hugging Face | The Space runs on Hugging Face's servers, on its shared ZeroGPU hardware | Hugging Face records your "IP address," session, and device, and keeps information "as long as necessary" (privacy policy, effective March 28, 2023). |
| Your own computer, in diffusers or ComfyUI | No one else | The weights sit on your disk and run on your own graphics card |
| Morse's Studio | Sent through Morse to generate the image | Studio images are stored in your browser, encrypted, not on Morse's servers; Morse doesn't train on them, sell them, or show ads (Morse privacy policy) |
The demo's code also has a prompt-rewriting option that would send prompts to Alibaba Cloud's DashScope API; as of October 23, 2026 it's labeled "DISABLED! Do not use." Hugging Face's policy doesn't mention prompts typed into Spaces.
Running it yourself is the one setup where no company sees the prompt, at the price of a capable graphics card, a download of 20 GB or more, and the setup. For more, read our guide to AI that doesn't save your data.
How to use Z-Image Turbo without a GPU, and keep your images on your device
The free ways to run Z-Image Turbo either send your prompt to a server under someone else's policy or ask you to build the server yourself. Studio gives you the model with no setup and keeps the results on your device. Studio, Morse's image generation, includes 3 images a day with Z-Image Turbo on the Free plan, with no credit card and no graphics card of your own. Your Studio images are stored in your browser, encrypted, not on Morse's servers, and Morse doesn't train AI on your content, sell your data, or show ads. Clearing your browser deletes them for good, and Morse can't recover them, so keep a copy of any you need.
Pro, at $19 a month ($17 billed yearly), adds unlimited use of Pro image models, including Nano Banana 2 Lite, Qwen Image 3 Pro, and 4 unfiltered image models. Unfiltered models are on paid plans only.
Tongyi's report calls the model "a powerful text decoder capable of translating detailed prompts into realistic images," so the prompt does much of the work. Draft it in a Morse chat first, in the privacy mode that fits:
- Private, to write the scene or the Chinese headline with an open model. GLM, Kimi, or DeepSeek drafts it on hosts that agreed to keep nothing (zero data retention).
- Anonymous, to have GPT, Claude, or Gemini tighten the prompt. The provider gets the draft without knowing it's you, and may keep it for a limited time to check for abuse, under its own policy.
- E2EE, on Pro, to plan a concept before you settle on a prompt. Work it out in a verified enclave (how it works), then paste the final prompt into Studio, which doesn't run in E2EE.
In every mode, your chat history lives only in your browser, encrypted and locked when you sign out, and Morse's servers don't keep messages or log message content.
Which way to run Z-Image Turbo
- You have a 16 GB card and want no one to see a prompt: run the weights in ComfyUI or diffusers, and add a safety filter if others will use it.
- You want one quick try: Tongyi's demo is free, under Hugging Face's policy.
- You want no setup, with your images kept on your device: use Studio on Morse, 3 images a day on Free.
Start free and make your first three Z-Image Turbo images today, with no card needed.
Common questions
Does Z-Image Turbo support negative prompts?
No. Tongyi's comparison table marks classifier-free guidance and negative prompting as unsupported in Turbo, and its sample code sets guidance_scale to 0. The base Z-Image model supports both, with a guidance scale of 3 to 5.
Can you train a LoRA on Z-Image Turbo?
People do. Hugging Face lists 880 adapters built on it as of October 23, 2026. Tongyi's own table rates Turbo's fine-tunability as N/A, though, and its card for the undistilled Z-Image calls that model "a good base for LoRA training."
What resolutions can Z-Image Turbo make?
Tongyi's technical report says the model can generate at arbitrary resolutions "up to the 1k-1.5k range." Its official demo offers presets built around 1024, 1280, and 1536 pixels, in aspect ratios from 21:9 to 9:21.
Who makes Z-Image Turbo?
The Z-Image Team at Tongyi Lab, part of Alibaba Group. It publishes under the Tongyi-MAI name on Hugging Face and GitHub.