Z-Image
Z-Image is a 6B single-stream diffusion transformer trained for roughly 314K H800 GPU-hours (about $630K) and competitive with much larger image models.
Scalable Single-Stream DiT (S3-DiT) with an 'omni' pretraining recipe supporting generation and editing, plus a distilled Z-Image-Turbo for fast inference on consumer hardware. Argues quality does not require ever-larger models.
- Date
- Thursday, 27 November 2025
- Lab
- Alibaba (Tongyi)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Training compute | about 314K H800 GPU-hours (~$630K) authors' estimate | authors |
| Parameters | 6B | authors |
Date is arXiv v1; later versions revised through 2026-07-06. Owning group (Tongyi) taken from the project's repo name, not the abstract page.
Sources
This record was checked against its sources on 6 October 2026. How we check