Z-Image is Tongyi-MAI's (Alibaba) 6B-parameter S3-DiT image model, released Nov 2025 under Apache 2.0 — fully commercial open weights. The Turbo variant generates in under a second on datacenter GPUs and runs on 16 GB consumer cards, renders bilingual English/Chinese text, and debuted #1 open-source on Artificial Analysis. Z-Image Turbo and Turbo V2 are available in the invideo agent.
Updated August 2026
Z-Image is a pointed argument against parameter-count maximalism. It is a 6-billion-parameter open-source image model from Tongyi-MAI, a lab under Alibaba's Tongyi ecosystem, whose distilled Turbo variant generates images in under a second on datacenter GPUs and runs on consumer cards with 16 GB of VRAM — and it debuted as the #1 open-source model on the Artificial Analysis text-to-image leaderboard (December 8, 2025). Released November 26, 2025 under the fully commercial Apache 2.0 license, it is also one of the few models of any size with strong bilingual English–Chinese text rendering. In the invideo agent's picker it appears as Z-Image Turbo and Z-Image Turbo V2.
What is Z-Image and who makes it?
Z-Image comes from Tongyi-MAI, which publishes the weights on Hugging Face and ModelScope. The pitch is efficiency without the usual quality tax: instead of scaling to tens of billions of parameters, the team built a compact 6B "single-stream DiT" — S3-DiT — in which text tokens, visual-semantic tokens, and VAE image tokens all flow through one unified transformer stream. A December 1, 2025 technical report on arXiv details the design, with companion papers on the Decoupled-DMD distillation method and DMDR reinforcement-learning refinement that produce the Turbo variant.
The result made it, at release, the fastest credible open-weight model in its quality class: 8 network function evaluations, sub-second latency on an H800, and a footprint small enough for a 16 GB consumer GPU — laptop-class hardware for near-frontier output.
Which Z-Image variants exist?
| Variant | Released | What it's for |
|---|---|---|
| Z-Image-Turbo | Nov 26, 2025 | Distilled 8-step model; the sub-second flagship for everyday generation |
| Z-Image (base) | Jan 27, 2026 | Undistilled 50-step CFG model; maximum quality and the fine-tuning starting point |
| Z-Image-Edit | Announced, unreleased at last check | Instruction-following image-to-image editing |
| Z-Image-Omni-Base | Announced, unreleased at last check | Raw pretrain checkpoint for generation + editing fine-tuning |
| Z-Image Turbo V2 (invideo picker) | — | The newer fast text-to-image + image-to-image variant available in the invideo roster |
A transparency note on that last row: as of this update, Tongyi-MAI's official Hugging Face org and GitHub model zoo list only Z-Image and Z-Image-Turbo. "Turbo V2" as listed in various tool rosters has no official standalone announcement we can cite, so this page describes it only by what it does in the picker — a faster variant that adds image-to-image alongside text-to-image — without attributing external specs to it.
Z-Image specs and hardware requirements
| Spec | Detail (as of August 2026) |
|---|---|
| Developer | Tongyi-MAI (Alibaba Tongyi ecosystem) |
| Parameters | 6B |
| Architecture | S3-DiT — single-stream DiT, unified text + visual-semantic + VAE token stream |
| Speed (Turbo) | 8 NFEs; sub-second on NVIDIA H800 |
| Min. consumer hardware | 16 GB VRAM GPU |
| Distillation / RL | Decoupled-DMD; DMDR |
| Text rendering | Bilingual — English and Chinese |
| License | Apache 2.0 (Turbo and base) |
| Leaderboard | 8th overall, #1 open-source, Artificial Analysis t2i (Dec 8, 2025) |
Every row points the same way — 6B parameters, 8 steps, 16 GB, Apache 2.0 all optimize for running cheaply anywhere; the December leaderboard rank is what proved quality survived the diet.
Sources: Hugging Face model card, GitHub README.
Where Z-Image wins — and where 6B shows
Strengths. The model card names three: photorealism, bilingual text rendering, and instruction adherence. The bilingual point is genuinely rare — Z-Image renders legible English and Chinese text in-image, which matters for anyone producing creative for both markets. The speed positioning is the other differentiator: sub-second datacenter latency and a 16 GB VRAM floor made it, per the December 2025 leaderboard result, the best open-source image model available at its debut while remaining self-hostable on a gaming PC.
Weaknesses. As a 6B model, it gives up some headroom to much larger frontier models on complex multi-subject scenes — it ranked 8th overall, not first, at debut, and the leaderboard has moved since December 2025. The official editing variant (Z-Image-Edit) remained unreleased at last check, so first-party editing is not yet part of the open-weights story. And as with any fast distilled model, the 8-step Turbo trades a little fine detail against the 50-step base.
How should you prompt Z-Image?
- Put exact text in quotes. For in-image text, spell it precisely —
a neon sign reading "OPEN LATE"— and keep it short; a few words render far more reliably than sentences, in English or Chinese. - Lean on instruction adherence. The model card highlights instruction following, so structured prompts (subject, setting, lighting, style, text) tend to land literally.
- Iterate fast, then refine. Use Turbo's speed for exploration; if you self-host and need maximum fidelity on a final frame, the 50-step base model is the quality pass.
What is Z-Image best used for?
- High-volume drafting and iteration — sub-second generations make it the natural first-pass engine inside an AI image generator workflow: explore ten compositions in the time a slower model renders one.
- Thumbnails with real text — bilingual in-image text rendering is exactly what an AI thumbnail maker workflow needs, where a title word baked into the image usually breaks on other models.
- Stylized and photoreal art on a budget — Apache 2.0 plus a 16 GB VRAM footprint means hobbyists and studios alike can run it for AI art without per-image fees when self-hosting.
Common Z-Image questions
Can I use Z-Image commercially?
Yes. Both Z-Image-Turbo and the Z-Image base model are licensed Apache 2.0, one of the most permissive licenses in the model ecosystem — commercial use, modification, and redistribution are all allowed, self-hosted or via a platform.
Is Z-Image really open source?
Fully, for the released models: weights are on Hugging Face and ModelScope under Apache 2.0, with a public technical report (arXiv 2511.22699, Dec 2025). The announced Edit and Omni-Base variants were not yet released at last check.
What hardware do I need to run Z-Image locally?
A GPU with 16 GB of VRAM, per the official model card — within reach of consumer cards. On datacenter hardware (H800), Turbo generates in under a second.
Is there an official Z-Image Turbo V2?
Tongyi-MAI's official channels list Z-Image and Z-Image-Turbo; no standalone "Turbo V2" release has an official announcement we can cite as of August 2026. Where "Z-Image Turbo V2" appears in tool rosters, including invideo's, treat it as the newer fast text-to-image + image-to-image option in that roster.
How good is Z-Image at text in images?
It's a signature strength: legible bilingual English and Chinese text rendering, called out on the official model card — unusual even among frontier-scale models.
Where can you use Z-Image without setting up a GPU?
Self-hosting is free but means drivers, VRAM, and setup. The zero-setup route: both Z-Image Turbo and Z-Image Turbo V2 are in the model roster at invideo, where the agent treats Z-Image as one of 200+ models it can sequence — generate a batch of sub-second stills, pick the keeper, and send it into a thumbnail or straight into a video shot without leaving the chat. The full roster lives on the AI models index; for image-first work, start from the AI image generator page.
Version history: first published August 2026, covering Z-Image-Turbo (Nov 2025), the base model (Jan 2026), and the announced Edit/Omni-Base variants.