What it is
Z-Image-Turbo was released on November 26, 2025, by Tongyi-MAI, an Alibaba research group. It is the distilled member of the 6-billion-parameter Z-Image family and generates an image in 8 function evaluations, with no classifier-free guidance. The model card claims sub-second generation on Nvidia H800 GPUs and says it fits on consumer cards with 16GB of VRAM. It is strongest at photorealism and at rendering English and Chinese text.
The weights use the Apache 2.0 licence, with no revenue threshold. The undistilled Z-Image base model followed on January 27, 2026. An editing variant, Z-Image-Edit, and a combined Omni base are still marked 'to be released' on the model card nearly a year after launch, and the GitHub repository has not been updated since February 2026.
Our review
Z-Image-Turbo earns its place through licence and efficiency rather than peak quality. Apache 2.0 with no revenue cap is still uncommon among strong image models, and eight steps in a 16GB footprint means it runs on a single gaming GPU. On fal it costs $0.005 per megapixel, so a standard 1024x1024 image is about half a cent.
Its ranking has slipped fast. In December 2025 the Z-Image team reported that Turbo stood eighth overall on Artificial Analysis's text-to-image board and first among open models. It now sits 71st of 105 at 936 Elo, against FLUX.2 [dev] at the board's fixed 1000 reference and Qwen-Image-2.1, the leading open model, at 1042. Its own undistilled base, Z-Image, ranks lower still at 891, which says the distillation and RL stage did real work.
The price of that distillation is variety. The model card rates Turbo's diversity as low and its fine-tunability as not applicable, so expect similar compositions across seeds and train LoRAs on the base model rather than on Turbo itself. The promised Z-Image-Edit has not appeared, which leaves teams that need editing to pair it with something else.
Against other cheap open options it still holds up. FLUX.2 [klein] 9B ranks just below it on the same board at 929 Elo while costing $15 per 1,000 images on its maker's API, about three times Turbo's price on fal.
Where it earns its place
An e-commerce marketplace generating lifestyle shots for 50,000 listings a month
Produce photorealistic product backdrops without per-seat design tools
Call fal's endpoint at $0.005 per megapixel, roughly $250 a month at 1024x1024. Low diversity helps keep a consistent house style but means prompts need explicit variation.
A machine-learning engineer with one RTX 4090
Self-host an image endpoint for an internal tool
Serve the weights through Diffusers on the 24GB card; the 16GB footprint leaves headroom. Apache 2.0 removes the licence review a revenue-capped model would trigger.
A Chinese-English bilingual marketing agency
Generate posters and social graphics with readable text in both scripts
Text rendering in both languages is the model's stated strength. Final layouts still need a designer, and anything needing edits must go to a separate editing model.
Before you commit
- Setup
- Install Diffusers from source and load Tongyi-MAI/Z-Image-Turbo; first images in minutes on a 16GB GPU, or call a hosted endpoint such as fal with no setup.
- Integrations
- Diffusers, vLLM-Omni and SGLang-Diffusion; ModelScope and Hugging Face weights; hosted on fal and the creator's API listed by Artificial Analysis.
- Data and privacy
- Apache 2.0 with no revenue threshold or acceptable-use annex; self-hosting keeps prompts and outputs local.
- Cost at scale
- On fal, 100,000 images a month at 1024x1024 is about $500, plus $0.0025 per request if prompt expansion is switched on; self-hosting needs one 16GB-plus GPU.
- Limits
- No negative-prompt or CFG control because guidance is distilled out; low seed diversity; no editing variant yet; LoRA training belongs on the base model.
What's good
- Apache 2.0 licence with no revenue cap
- Eight-step generation that fits in 16GB of VRAM
- Strong bilingual English and Chinese text rendering
- About half a cent per 1024x1024 image on fal
What's not
- 71st of 105 on Artificial Analysis's text-to-image board
- Low output diversity, by the model card's own rating
- Fine-tuning has to go through the slower base model
- Edit and Omni variants still unreleased
Best for
Developers who need a commercially unrestricted open image model that runs cheaply on consumer GPUs or for fractions of a cent per image.
Alternatives
Z-Image-Turbo: common questions
- Is Z-Image-Turbo worth it?
- Z-Image-Turbo is what you run when the licence and the GPU bill matter more than placing in the top half of the leaderboard. We rate it 3.7 out of 5.
- How much does Z-Image-Turbo cost?
- Free weights under Apache 2.0. fal charges $0.005 per megapixel; Alibaba Cloud Model Studio $0.015 per image ($0.03 with prompt rewriting).
- What are the downsides of Z-Image-Turbo?
- 71st of 105 on Artificial Analysis's text-to-image board; low output diversity, by the model card's own rating; fine-tuning has to go through the slower base model; edit and Omni variants still unreleased.
- Who is Z-Image-Turbo best for?
- Developers who need a commercially unrestricted open image model that runs cheaply on consumer GPUs or for fractions of a cent per image.
- How much does Z-Image-Turbo cost for a team or in production?
- On fal, 100,000 images a month at 1024x1024 is about $500, plus $0.0025 per request if prompt expansion is switched on; self-hosting needs one 16GB-plus GPU.
- Is Z-Image-Turbo safe for company data?
- Apache 2.0 with no revenue threshold or acceptable-use annex; self-hosting keeps prompts and outputs local.
- What are the best Z-Image-Turbo alternatives?
- The closest alternatives we have reviewed are FLUX 3 Image Klein, Qwen-Image-2.1, Ideogram 4.0 and Nano Banana 2.1.
Answers checked Oct 23, 2026.