Z-Image Turbo

Z-Image Turbo generates images from a text prompt only, producing photorealistic results in under a second of inference time thanks to its low step count — there's no editing or reference-image mode on this model.

On AutorunX it's positioned as the fastest, cheapest image option on the platform, used for lightweight content like banner ad variations and covers where speed and volume matter more than the last increment of fidelity.

Z-Image Turbo is built on a Scalable Single-Stream DiT (S3-DiT) architecture, where text tokens, semantic vision tokens, and VAE image tokens are all concatenated into one unified input stream rather than processed through separate branches. That single-stream design, combined with a distillation technique called Decoupled Distribution Matching Distillation, is what lets it generate a finished image in as few as 8 inference steps — a fraction of what a typical diffusion model needs.

The model is a compact 6-billion-parameter design, small next to most current-generation image models, yet it still ranked as the top open-source model on independent text-to-image leaderboards at release — evidence that the distillation approach preserves most of the quality a larger, slower model would produce.

It also carries native bilingual text rendering, handling complex Chinese typography, English text, and mixed-language layouts in the same generation, a capability that's inherited from Tongyi Lab's broader Qwen-family training approach. On AutorunX, Z-Image Turbo is accessed via a hosted cloud API rather than self-hosted on AutorunX's own GPU fleet, distinguishing its cloud track from the owned FLUX.2 Klein and Qwen-Image-Edit models.