Z-Image is the text-to-image model Alibaba's Tongyi-MAI team open-sourced in November 2025: six billion parameters, Apache 2.0 weights, and a distilled Turbo build that reaches a finished image in eight sampling steps. It occupies the bottom rung of the price list on BulkImagen, which is what makes a genuinely long bulk z-image run worth setting up.
One per line — each line is a separate prompt.
Z-Image generates from text prompts only, so there's no reference-image upload.
Z-Image is not the largest model BulkImagen carries and is not trying to be. Its entire argument is ratio — how much a small network gives back for the compute it asks for — which is a different argument from most of the models here.
Artificial Analysis put Z-Image Turbo at the top of its Image Arena among open-weight text-to-image models when it appeared, ahead of FLUX.2 [dev] and Qwen-Image. The Z-Image technical report makes the same case from the other direction: an 87.4% G+S Rate on roughly a fifth of FLUX.2 dev's 32B parameter count.
Text is where the Z-Image paper puts its sharpest results — 0.8671 word accuracy on CVTG-2K, and 0.935 English against 0.936 Chinese on LongText-Bench, reported ahead of GPT-Image-1 and Qwen-Image. If the batch carries captions in either language, that benchmark is the reason to look at Z-Image at all.
The Z-Image model card and the paper both lead with photorealism, and the distilled Turbo weights get there in eight sampling steps rather than the usual dozens. The open release fits on a 16GB card, but nothing on this page asks you to own one — the batch runs on our side.
Aspect ratios, the resolution tier, prompt length and the credit rate for Z-Image, pulled out of the running configuration rather than written into a paragraph where they would slowly go out of date.
Z-Image reads as having no reference-image support on purpose — the FAQ below explains what that means before you plan a batch around it.
Z-Image reads words and nothing else, so this workflow is shorter than the one on the reference-driven model pages. Describe, frame, collect.
There is no image input, so any detail you want has to live in the prompt itself. The Z-Image paper is also candid that six billion parameters buy limited world knowledge, so describe a niche subject rather than assuming the model recognises the name.
Aspect ratio and how many images each prompt returns — that is the whole form, because this page is pinned to Z-Image and it exposes one resolution tier. The credit total appears on the confirmation screen before the first call goes out.
Results arrive as they finish. The Z-Image Turbo model card rates its output diversity as low, so if rows come back near-identical the prompts were probably too close together — rewrite the wording rather than re-running the same line and hoping.
A small, fast, text-strong model has a shape to it, and Z-Image is no exception. These are the jobs that fit that shape instead of fighting it.
Layouts carrying English or Chinese lettering are the case the Z-Image benchmarks were built around. Produce a whole set of screens or poster variants in one pass, then read the type on each before committing to any of them.
Photorealism is what Tongyi-MAI foregrounds about Z-Image, and quantity is what this page adds. Lifestyle frames, stock-style scenes and background plates come back as a set you can sift, not as one considered render at a time.
If you have Z-Image checked out locally, the same family is available here without a GPU queue in front of it. Draft and sort a prompt set against the hosted Turbo build, then take the shortlist back to your own environment.
Anything else? Email support@bulkimagen.com
The cheapest slot on the platform, a Z-Image build with published bilingual text scores, and no reference uploads to prepare first. Paste the prompts, run them as one batch, and take the ZIP away.
