Z-Image Turbo — Open-Source 6B Text-to-Image from Alibaba Tongyi Lab
Introducing Z-Image Turbo
Z-Image Turbo is a 6-billion-parameter text-to-image model from Alibaba's Tongyi Lab — the team behind Qwen — released as open source on November 26, 2025 under the Apache 2.0 license. Distilled to just eight sampling steps via Decoupled-DMD, it generates high-quality images in seconds, delivers native bilingual Chinese and English text rendering, and ranks #1 among open-source image models on the Artificial Analysis leaderboard.
Z-Image Turbo is built on a Scalable Single-Stream DiT (S3-DiT) architecture, in which text, semantic vision tokens, and VAE image tokens are concatenated into a unified input stream — a design Tongyi Lab uses to maximize parameter efficiency at the 6B scale. The base Z-Image model is distilled into Z-Image Turbo via Decoupled-DMD, collapsing inference to eight function evaluations, and aligned to human aesthetic preference with DPO and GRPO. The result is sub-second latency on data-center GPUs and comfortable inference on consumer cards with 16 GB of VRAM.
Native bilingual text rendering is the model's strongest differentiator. Z-Image Turbo handles complex Chinese typography — signage, posters, packaging — alongside English text in the same image, a capability most Western image models still struggle with. As of early 2026, Z-Image Turbo holds the #1 position among open-source image models on the Artificial Analysis Text-to-Image Leaderboard and the top open-source slot on Alibaba AI Arena, with weights freely available on Hugging Face and ModelScope under Apache 2.0 (commercial use permitted).
On LoveGen AI, Z-Image Turbo accepts prompts up to 2000 characters and offers nine preset aspect ratios — 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 2:1, and 1:2 — alongside custom dimensions in the 376–1536 pixel range. A deterministic seed parameter (1 to 2,147,483,647) makes outputs reproducible for iteration and batch work. End-to-end generation typically completes in around ten seconds for 0.1 credit per image, making Z-Image Turbo our most cost-efficient text-to-image option — well-suited to high-volume social content, Chinese-language creative work, and rapid concept exploration. Generated image URLs remain valid for 24 hours.
How to Use Z-Image Turbo
Write Your Prompt
Describe the image you want in up to 2000 characters. Be specific about subject, style, lighting, and composition for the best results.
Pick an Aspect Ratio
Choose one of nine preset aspect ratios that fits your destination — square for social, 16:9 for thumbnails, 9:16 for vertical video covers.
Generate & Save
Click Generate. Your image arrives in roughly ten seconds. Download it within 24 hours since the generated link expires after that.
Z-Image Turbo Technical Specifications
| Developer | Alibaba Tongyi Lab (Tongyi-MAI) |
| Release Date | November 26, 2025 |
| License | Apache 2.0 (open-source, commercial use permitted) |
| Architecture | Scalable Single-Stream DiT (S3-DiT) |
| Parameters | 6 billion |
| Inference Steps | 8 (distilled via Decoupled-DMD) |
| Mode | Text-to-image |
| Native Languages | Chinese + English text rendering |
| Estimated Generation Time | ~10 seconds end-to-end |
| Prompt Length | Up to 2000 characters |
| Aspect Ratios | 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 2:1, 1:2 |
| Custom Dimensions | 376–1536 px (width × height) |
| Reproducibility | Seed parameter (1 to 2,147,483,647) |
| Content Moderation | Always-on baseline + strict NSFW filter |
| Output Validity | 24 hours (save outputs promptly) |
| Cost | 0.1 credit per image |
Why Choose Z-Image Turbo
Native Bilingual Text Rendering
Accurate Chinese and English typography in the same image — including signage, posters, and packaging — a capability most Western image models still lack.
Open-Source by Alibaba Tongyi Lab
Built by the team behind Qwen and released under Apache 2.0 in November 2025. Top-ranked open-source image model on the Artificial Analysis leaderboard.
Distilled 6B Architecture
A Single-Stream DiT (S3-DiT) with 6 billion parameters, distilled to just 8 sampling steps via Decoupled-DMD for sub-second inference on GPU.
Lowest Cost per Image
0.1 credit per generation — LoveGen AI's most efficient text-to-image option for high-volume work.
Reproducible with Seeds
A deterministic seed parameter locks in results. Same prompt plus same seed produces consistent output across runs.
Z-Image Turbo vs Other AI Image Generators
| Feature | Z-Image Turbo | GPT Image 2 | Flux 2 Pro | Ideogram v3 |
|---|---|---|---|---|
| Developer | Alibaba Tongyi Lab | OpenAI | Black Forest Labs | Ideogram |
| License | Apache 2.0 (open-source) | Closed | Closed | Closed |
| Parameters | 6B | Undisclosed | Undisclosed | Undisclosed |
| Primary Strength | Bilingual text + open-source | Multi-image editing | Studio quality | Typography & branding |
| Generation Time | ~10 seconds | ~30 seconds | ~30 seconds | ~15 seconds |
| Aspect Ratios | 9 presets + custom | 3 presets + auto | Multiple | Multiple |
| Custom Dimensions | Yes (376–1536 px) | No | Yes | Limited |
| Image Input | No | Up to 4 images | Up to 8 images | No |
| Cost per Image | 0.1 credit | Higher | Higher | Higher |
| Best For | Bilingual content & fast iteration | Editing & blending | Studio work | Logos & posters |
Popular Uses for Z-Image Turbo
Rapid Concept Exploration
Generate many variations quickly to explore visual directions for branding, campaigns, or product ideas.
Social Media Content at Scale
Produce posts, stories, and ad creatives in any aspect ratio at low per-image cost for high-volume content schedules.
Thumbnails & Banners
Use 16:9 and 9:16 presets for video thumbnails and vertical covers, or custom dimensions for site banners.
Explore Related AI Image Generators
GPT Image 2
OpenAI's image model with multi-image reference editing and natural-language prompts.
Nano Banana Pro
Google's image model with up to 14-image blending and Gemini-class prompt understanding.
Flux 2 Pro
Black Forest Labs' studio-grade generator with 4MP resolution.
Ideogram v3
Industry-leading typography and text rendering for logos and posters.
Qwen Image
Alibaba's sister image model from the Qwen family, with strong multilingual prompt understanding.
Midjourney V7
Industry-leading aesthetic image generation that returns four candidates per task.
Frequently Asked Questions About Z-Image Turbo
Who built Z-Image Turbo?
Z-Image Turbo was developed by Alibaba's Tongyi Lab — the same team behind the Qwen model family — and released as open source under the Apache 2.0 license on November 26, 2025.
How fast is Z-Image Turbo?
Z-Image Turbo is distilled to just 8 sampling steps via Decoupled-DMD, giving sub-second inference on data-center GPUs. End-to-end on LoveGen AI, generation typically completes in around ten seconds.
Can Z-Image Turbo render Chinese and English text?
Yes — native bilingual text rendering is one of Z-Image Turbo's biggest differentiators. The model handles complex Chinese typography, English text, and mixed-language layouts that many Western image models still struggle with.
What aspect ratios does Z-Image Turbo support?
On LoveGen AI, Z-Image Turbo offers nine preset aspect ratios — 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 2:1, and 1:2 — and custom dimensions in the 376–1536 pixel range.
Can I reproduce a specific image with Z-Image Turbo?
Yes. Z-Image Turbo accepts a numeric seed parameter (1 to 2,147,483,647). The same prompt with the same seed produces consistent results, useful for iterating or creating series of related images.
What does Z-Image Turbo cost on LoveGen AI?
Z-Image Turbo costs 0.1 credit per generated image — our most cost-efficient text-to-image model. Generated image URLs remain valid for 24 hours, so download outputs promptly.