whatmodel.sh / image / Z-Image-Turbo
Z-Image-Turbo
↗ share
+ contribute
Z-Image-Turbo is an open-weights image model from Alibaba (Tongyi-MAI), released November 2025. Its published size is 6B parameters. Licence: apache-2.0.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of Z-Image-Turbo, out of 197 with a score, then more from Alibaba (Tongyi-MAI).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 76 |  Qwen-Image-2512 | 56±1 | api $0.020/image +local | ||
| 77 |  Reve V1.1 Fast | 56±4 | comparison only | ||
| 78 | Z-Image-Turbo | 55±1 | api $0.0050/image +local | ||
| 79 |  LLaDA-Image | 55±3 | local | ||
| 80 | Qwen-Image-Edit | 55±2 | api $0.030/image +local | ||
| more from Alibaba (Tongyi-MAI) | |||||
| 63 | Z-Image (Base) | 58±2 | api $0.010/image +local | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #55 of 126 · 95% of its score · 17 tests
Beats HunyuanImage 2.1 here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LMArena text-to-image | 1083.5 ✎ | ▂▂▃▄▇█▄▂▄▂·▂ beats 22/75 | 30% |
| QC-T2I-Bench (English, HCQ total) | 85.7% ✎ | ▄····▄█▄▄█▄█ beats 4/10 | 10% |
| UltraText Bench (Composite, bilingual dense text rendering) | 54.4% ✎ | ▁▂▂█▃▁▂▃·▁·▁ beats 12/22 | 8.0% |
| ImageTime (prompt-only, overall) | 54.1% ✎ | ▄·····▄█▄·▄█ beats 3/7 | 6.0% |
| LongText-Bench (EN) lab | 91.7% ✎ | ▃·····▅██▅▃▃ beats 4/12 | 5.9% |
| LongText-Bench (CN) lab | 92.6% ✎ | ▃·······████ beats 6/12 | 4.6% |
| GenEval lab | 82.0% ✎ | ▇▃▅▇▁▄▅▃█▃▃▅ beats 29/39 | 4.3% |
| FEPBench instruction faithfulness | 39.0% ✎ | ▄·█···▄▄▄▄▄▄ beats 1/8 | 3.9% |
9 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CVTG-2K word accuracy lab | 85.9% ✎ | ▄▄▄█▄▄·▄···▄ beats 2/8 | 3.6% |
| OneIG-Bench (CN) lab | 50.7% ✎ | ▃·▃█▃·▃··▃·▃ beats 3/8 | 3.5% |
| OneIG-Bench (EN) lab | 52.8% ✎ | ▃······▃▅█▅▃ beats 2/9 | 3.5% |
| DPG-Bench lab | 84.9% ✎ | ▄·▂·▂▆█▄▂▄▂▂ beats 5/17 | 3.2% |
| TIIF-Bench testmini (long prompts) lab | 80.0% ✎ | ▄······█··▄▄ beats 0/4 | 2.8% |
| TIIF-Bench testmini (short prompts) lab | 77.7% ✎ | █···█···██·█ beats 0/4 | 2.8% |
| HPSv3 score of Cosmos 3 report text-to-image evaluation images | 11.3600 ✎ | ▄···█··▄·▄·▄ beats 1/5 | 1.8% |
| CVTG-102ch PNED (Chinese character rendering, Cosmos 3 report) | 73.3% ✎ | ▄·······▄█▄▄ beats 4/5 | 0.5% |
| CVTG-500L PNED (English long-prompt text rendering, Cosmos 3 report) | 87.0% ✎ | █······█████ beats 2/5 | 0.0% |
other tests1 not in a task · 4.4% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| UniGenBench + Physical-AI prompts, All 1,170 (Cosmos 3 report) | 78.1% ✎ | ▄···█··▄···█ beats 0/5 | 4.4% |
$ where to run it
Through an API
api · Replicate (PrunaAI optimized) · $0.0050/image · replicate.com ↗ · price sourceUSD per output image up to 1 MP (1,000,000 pixels) on Replicate, as stated on the model page (target resolution 1; billed $5 per 1,000 output megapixels). Up to 0.5 MP costs $0.0025 and up to 2 MP $0.01. PrunaAI optimized serving. Product: Replicate model API. API model: prunaai/z-image-turbo. Settings: target resolution=1 MP, one output image. Price basis: published list rate. Currency: USD. Replicate official model page, observed 2026-09-25.
Other ways to run it
version
- Run it as published · needs ~7.7 GB · from github.com/leejet
sd-cli --diffusion-model models/z_image_turbo-Q3_K.gguf --vae models/ae.safetensors --llm models/Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p 'a cat holding a sign' --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 1024 --steps 8
Official weights and documentation: huggingface.co/Tongyi-MAI ↗
$ how the score was worked out · Tongyi-MAI (Alibaba) — Z-Imag… 34%, LMArena leaderboard dataset 30%, QC-T2I-Bench: Scalable Questi… 10% and 4 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json