Z-Image Prompting Guide

Z-Image (造相) is an open-weights image generation family from Alibaba's Tongyi-MAI, Apache-2.0, 6B parameters, built on a Scalable Single-Stream DiT (S3-DiT) architecture. It runs locally: the Turbo variant fits in 16 GB VRAM and reaches sub-second latency on an H800.

Facts here are verified against Tongyi-MAI's own sources (GitHub README, Hugging Face model card, project blog) on 2026-08-05, with the model family re-checked and mflux behaviour updated from its release notes on 2026-11-06 (no new Z-Image weights since January 2026). Community-sourced claims are labelled as such. See "Verification & Sources".

Read this first: the Turbo / base fork decides how you prompt

This is the single most important thing about Z-Image and the most common source of frustration. Turbo and the base model are prompted differently, and Turbo removes two things people assume every diffusion model has.