Imagifly — Z-Image Turbo

Z-Image-Turbo is the distilled text-to-image model the Z-Image team at Tongyi MAI (Alibaba Group) released on 26 November 2025. The card describes a 6B model; the published weights index at 6.15B in F32 and ship under Apache-2.0.

  • Text to Image
  • Fast

From 2 credits per image

Fashion street portrait sample for Z-Image Turbo
Sample output: Fashion street portrait sample for Z-Image Turbo

About Z-Image Turbo

Z-Image-Turbo is the distilled text-to-image model the Z-Image team at Tongyi MAI (Alibaba Group) released on 26 November 2025. The card describes a 6B model; the published weights index at 6.15B in F32 and ship under Apache-2.0.

It uses a Scalable Single-Stream DiT (S3-DiT), concatenating text tokens, visual semantic tokens and VAE tokens into one sequence instead of running two streams. Turbo is distilled from Z-Image and needs only eight forward passes (num_inference_steps=9, which resolves to eight DiT forwards) with CFG switched off (guidance_scale=0.0).

The low step count is the selling point: the team reports sub-second latency on H800 GPUs and that the model fits within 16GB of VRAM on consumer cards. Those figures are vendor claims; the sample image on the card is 1024x1024.

What it is good at

  • Eight steps per image with CFG disabled
  • 6B-class model that runs inside 16GB of VRAM
  • Official samples use 1024x1024
  • Renders Chinese and English text inside the image

Technical specs

Developer
Z-Image Team, Tongyi MAI (Alibaba Group)
Released
2025-11-26
License
Apache-2.0
Architecture
Scalable Single-Stream DiT (S3-DiT)
Parameters
6B (published weights index at 6.15B, F32)
Task
Text-to-image
Recommended settings
Documented defaults: num_inference_steps=9 (eight DiT forwards) and guidance_scale=0.0. No usable resolution range is published for the Turbo model.

Source: Tongyi-MAI/Z-Image-Turbo model card