# Z-Image > High-quality AI image generation model - Canonical page: https://artifio.ai/models/z-image - Run it: https://artifio.ai/create/z-image - Modality: prompt to image - Model brand: alibaba - Pricing: from about $0.005 per generation (exact cost shown in the workspace before you confirm; varies with selected options) - Typical generation time: ~30s (catalogue estimate) - Full pricing table: https://artifio.ai/pricing/models ## About Z-Image Z-Image is the 6B-parameter image generation foundation model from Tongyi-MAI, Alibaba's Tongyi lab. Its technical report (November 2025) describes a Scalable Single-Stream Diffusion Transformer (S3-DiT) that concatenates text tokens, visual semantic tokens, and image VAE tokens into one input stream, an efficiency-first design aimed squarely at the 20B-plus models that dominated open image generation. The whole model trained in roughly 314,000 H800 GPU hours, a fraction of typical foundation-model budgets, and the weights are Apache 2.0. This entry runs the base (undistilled) checkpoint, released on Hugging Face and ModelScope on January 27, 2026 after the Turbo variant had already shipped. The base model is the full-capacity member of the family: classifier-free guidance, robust negative prompting, high output diversity, and fine-tuning support that the distilled Turbo gives up. On Artifio it runs pay-per-generation in the same workspace as 100+ other models. ## Features - 6B-parameter S3-DiT architecture with a single unified input stream for text, semantic, and VAE tokens - Full sampling control: classifier-free guidance (recommended scale 3.0 to 5.0) and negative prompting, at 28 to 50 inference steps - Resolution range from 512x512 up to 2048x2048 - Photorealistic generation plus cinematic digital art, anime, and stylized illustration from one checkpoint - Bilingual text rendering in English and Chinese - Apache 2.0 weights, with the family also spanning Z-Image-Turbo (distilled), Z-Image-Edit (instruction editing), and Z-Image-Omni-Base ## What people use it for - Prompt-heavy creative work that benefits from negative prompts and guidance-scale control - Generating varied candidate sets, since the base model keeps the diversity the Turbo distillation trades away - Photorealistic portraits and scenes with embedded English or Chinese text - Base checkpoint for teams that fine-tune styles on their own data - Large-format 2K renders for print and web hero images ## Reported strengths - The technical report claims results comparable to or surpassing leading competitors, including commercial systems, at a 6B size that undercuts 20B to 80B rivals - High output diversity and easy fine-tunability relative to Z-Image-Turbo, per the family's own comparison table - Documented low training cost (about 314K H800 GPU hours), which backs the efficiency claims with numbers - Open Apache 2.0 license across the family ## Reported limitations - At 28 to 50 steps the base model is several times slower per image than the 8-step Turbo variant - The base model card publishes little benchmark data of its own; the family's headline leaderboard results were earned by Z-Image-Turbo - Instruction-based editing is not part of this checkpoint; that job belongs to the separate Z-Image-Edit variant - Local use still expects a 16GB-VRAM-class GPU unless CPU offloading is enabled ## Sources 1. Tongyi-MAI/Z-Image model card (Hugging Face (Tongyi-MAI)): https://huggingface.co/Tongyi-MAI/Z-Image 2. Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer (arXiv): https://arxiv.org/abs/2511.22699 3. Tongyi-MAI/Z-Image repository (GitHub (Tongyi-MAI)): https://github.com/Tongyi-MAI/Z-Image --- About Artifio.ai: 100+ AI image, video, and audio models in one workspace with a single shared wallet — no per-model subscriptions. This is the markdown rendition of https://artifio.ai/models/z-image for AI assistants and crawlers; the HTML page is canonical.