Z-Image: Powerful and highly efficient image generation model with 6B parameters (github.com)

Z-Image is a new 6B-parameter image-generation family from Tongyi Lab that prioritizes speed and efficiency while delivering photorealistic quality and strong bilingual (Chinese/English) text rendering. There are three public variants: Z-Image-Turbo (an 8-NFE distilled model that achieves sub-second inference on enterprise H800 GPUs and runs within 16GB VRAM consumer cards), Z-Image-Base (the undistilled checkpoint for community fine-tuning), and Z-Image-Edit (fine-tuned for instruction-driven image editing). Z-Image-Turbo reportedly matches or exceeds leading competitors in Elo-style human preference tests on Alibaba AI Arena and is provided as a diffusers pipeline (Tongyi-MAI/Z-Image-Turbo) — authors recommend bfloat16, optional Flash Attention, guidance_scale=0 for Turbo, and installing diffusers from source via the merged PRs. Technically, Z-Image uses a Scalable Single-Stream DiT (S3-DiT) architecture that concatenates text tokens, visual semantic tokens, and VAE tokens into a single sequence for better parameter efficiency than dual-stream designs. Few-step performance is enabled by Decoupled-DMD, a distillation approach that separates CFG Augmentation (the main driver) from Distribution Matching (a regularizer), and by DMDR, which integrates Reinforcement Learning with DMD during post-training to boost semantic alignment, high-frequency detail, and coherence. Cache-DiT support (DBCache, context/tensor parallelism) further accelerates inference, making Z-Image immediately useful for researchers and practitioners needing fast, editable, high-quality open models.