Z-Image Turbo: the 6B open model that actually renders text
Written by Clement
Our verdict
The best open-weight model available for images that contain text, and the one most likely to hand your subject a broken finger.
Reach for it when
- Posters, packaging, signage and anything where words have to be spelled correctly
- People on a 16GB consumer card who do not want to quantise
- Bilingual work: it handles Chinese and English typography in the same frame
- Commercial projects that need a permissive licence (Apache 2.0)
Skip it if
- Your work is close-up hands, jewellery or anything where finger anatomy is the subject
- You need a single character to stay identical across a set without training a LoRA
- You want built-in editing or inpainting: this is text-to-image only
At a glance
- Made by
- Alibaba Tongyi Lab, released 27 Nov 2025
- Size
- 6B parameters, S3-DiT architecture
- Speed
- 8 sampling steps, sub-second on data-centre GPUs
- Runs on
- 16GB VRAM consumer cards, ComfyUI
- Licence
- Apache 2.0, commercial use permitted
- Ranking
- #1 open-weights on the Artificial Analysis Image Arena
Alibaba's Tongyi Lab put Z-Image Turbo out on 27 November 2025 under Apache 2.0, and the number that got attention was 6 billion parameters. That is small. Flux 2 dev is several times the size. Small usually means worse, and here it did not: Artificial Analysis put Z-Image Turbo at the top of its open-weights image ranking, above Flux 2 dev, HunyuanImage 3.0 and Qwen-Image.
It gets there through distillation rather than scale. The base model is compressed into Turbo using Decoupled-DMD, which collapses generation down to eight sampling steps, then aligned to human aesthetic preference with DPO and GRPO. Eight steps on a 16GB card is the practical headline. You do not need a quantised build and you do not need to wait.
We ran it through the same three prompts we run every image model through. It did one thing no other model in the set managed, and one thing badly enough that it changes what you should use it for.
The text rendering is the reason to install it
Most open image models treat text as texture. Ask for a word and you get letter-shaped marks that dissolve under inspection. Z-Image Turbo spells.
Our first showdown prompt asks for a small tattoo in fine linework on a subject's forehead, reading "GenLovers". Krea 2, which beats Z-Image on almost every anatomy check, simply declined to render it. Z-Image Turbo produced it legibly, in the right place, following the curve of the brow. That is a hard ask: small text, on skin, on a curved surface, at an angle.
The published capability matches what we saw. Tongyi built it to handle Chinese typography in signage, posters and packaging alongside English in the same image, which is a harder problem than either alone and something most Western models still fumble. If your output has words in it, this is the open model to reach for.
Full specification
| Architecture | Scalable Single-Stream DiT (S3-DiT): text, visual semantic tokens and image VAE tokens concatenated into one input stream |
|---|---|
| Distillation | Decoupled-DMD, collapsing inference to eight function evaluations |
| Preference alignment | DPO and GRPO against human aesthetic preference |
| Parameters | 6B |
| Minimum VRAM | 16GB for the full-precision build |
| Sampling steps | 8 |
| Licence | Apache 2.0 |
| Editing / inpainting | Not included. Text-to-image only out of the box |
| Text rendering | Chinese and English, including mixed-script layouts |
How it compares to Krea 2
These two models came out of our showdown as a clean pair of opposites, which is more useful than a ranking.
Krea 2 is the better photographer. Hands, skin, fabric and light are stronger, and it produced the single best frame in our whole set. It also ignored two explicit instructions to get there: no tattoo, and a composition adjacent to the one we asked for.
Z-Image Turbo is the better listener. It rendered the text, kept two mirror reflections tracking the subject, and produced three genuinely distinct character designs where Krea 2 cloned faces. Then it broke a hand.
So the choice is not which model is better. It is whether your image lives or dies on anatomy or on instruction-following. Portrait and product work: Krea 2. Posters, signage, multi-character scenes and anything with words: Z-Image Turbo.
The model showdown
Every image model on this site runs the same six prompts across three tiers — SFW, erotic and explicit — on the same rig, with no cherry-picking and no retouching. Below is what Z-Image Turbo returned, graded against a fixed rubric so the results mean the same thing from one model page to the next.
- Run on
- Z-Image Turbo, FP8, ComfyUI on a single RTX 4090, 8 steps
- Date
Why these six prompts? ▾Why these six prompts? ▴
They are chosen to break things, across three tiers so a model that refuses NSFW outright still gets a fair, comparable test. The SFW tier stacks two-person contact, readable logo text on folded fabric and genuine expressions with teeth. The erotic tier stacks the four failure modes that mark an image as machine-made at a glance — hands, fine text on skin, a mirror that has to obey geometry, and three distinct character designs in one frame. The explicit tier exists to test the same anatomical limits at full intensity, but its output is never published on this site.
No model passes every check, and not every model can attempt every tier. That is the point. A rubric everything passes tells you nothing about which model to reach for on a Tuesday afternoon.
SFW detail test
Not run yet
Re-run queued. The first pass showed more skin than this site publishes (topless/underwear-only) — past the line drawn for the ranking domain, so it was withheld rather than published. Queued to re-run fully clothed.
Erotic tier
Solo portrait, low light
Skin under warm low-key light, lace texture, hand and ring anatomy, fine tattoo linework, and a mirror reflection that has to agree with the subject. The single hardest frame of the set.
Scorecard
- Partial: Hands and fingers anatomically correct
- Pass: Fine tattoo linework legible
- Pass: Mirror reflection consistent with the subject
- Pass: Lace and fabric texture holds up close
What we saw
This is the only model in the set that rendered the forehead tattoo, and it rendered it legibly: "GenLovers" in fine linework, correctly placed, following the curve of the brow. It also produced two mirror reflections that genuinely track the subject's pose. Both are hard. The cost shows in the hands: the supporting hand has an extra knuckle break and the fingers taper wrong. Z-Image spends its budget on the thing you asked for and skimps on the thing you assumed.
Three characters, anime, night
Stylized rendering with three distinct character designs in one frame, water and reflections, bare-foot anatomy, and two animals. Multi-subject scenes are where most image models quietly give up.
Scorecard
- Pass: Three distinct, non-cloned character designs
- Partial: Feet and toes rendered correctly
- Pass: Water surface and reflections coherent
- Partial: Both animals recognizable and correctly anatomized
What we saw
Three distinct designs: different hair, different skin tones, different builds, no face-cloning. That is the check most models fail and Z-Image passes cleanly. Water and reflections are strong again. Two problems: the feet under the water surface lose toe definition, and the prompt asked for one cat and one dog and got one cat and two dogs. Counting is not a diffusion model's strength, and Z-Image is not the exception.
Explicit tier
Not run yet
Queued. Waiting on the two explicit-tier prompts being authored (docs/MODEL_SHOWDOWN_FRAMEWORK.md), then a human-confirmed run posted to the private Discord channel.
Frequently asked questions
- Is Z-Image Turbo free?
- Yes, genuinely. The weights are published under Apache 2.0, which permits commercial use, and you run them on your own hardware, so there is no per-image cost and nothing you generate touches anyone else's servers. Your costs are the GPU and the electricity. That is a meaningfully different deal from "free tier" models that meter you.
- What GPU do you need to run Z-Image Turbo?
- 16GB of VRAM runs the full-precision build comfortably, which puts it in reach of a 4060 Ti 16GB, a 4080 or a 3090. That is the design goal: Tongyi built a 6B model specifically so it would fit consumer cards without quantisation. On data-centre hardware the eight-step generation is sub-second.
- Is Z-Image Turbo better than Flux 2?
- On the Artificial Analysis Image Arena it ranks above Flux 2 dev among open-weights models, and it is roughly a quarter of the size. In our own testing the honest answer is narrower: it is better at text and at following an unusual instruction, and it is not better at hands. Benchmarks aggregate across prompt types, so treat the ranking as a reason to try it rather than a verdict on your particular work.
- Can Z-Image Turbo edit an existing image?
- Not out of the box. It ships as text-to-image only, with no inpainting or instruction-based editing, which is the same gap Krea 2 launched with. If you need to change one element of an image you already have, you will be pairing it with a separate editing model or doing the pass in ComfyUI yourself.
- Why does Z-Image Turbo only need 8 steps?
- Because Turbo is a distilled model, not the base one. Tongyi compressed the base Z-Image using Decoupled-DMD, which trains the small model to reproduce in eight function evaluations what the base model does across many more. You are trading a little diversity for a large speed gain, which is the same bargain Krea 2's Turbo variant makes.
Hands-on guides
Related models
Get new guides by email
One email when we publish new guides and model breakdowns. No spam, unsubscribe anytime.