Z-Image-Turbo came out on Nov 27th 2025 and it lets you generate an image from plain language prompts. It's said to follow instructions well.
The model comes from a group called Tongyi-MAI which seems is part of Alibaba’s AI setup, likely linked to the wider Tongyi model family that started inside Alibaba Cloud.
Some info from FAL suggests the tool runs fast, with about a one-second delay and early examples look pretty promising especially considering model's tiny size.
The model has 6 billion parameters and was built for speed. It’s been distilled with 8 NFEs which helps it run in under a second on strong GPUs, and it works fine on setups with 16 GB of VRAM.
Tongyi-MAI has now put Z-Image-Turbo on Hugging Face. It comes with an Apache-2.0 license, so you can download it, use it, and even use it for business stuff without trouble.
It’s made for high-quality photorealistic image creation. It supports both English and Chinese text drawing, and it follows prompts closely. So it’s a strong and fast tool for developers or creators who want a text-to-image model that’s easy to use.
Z-Image comes in 3 versions:
Z-Image-Turbo - the lighter, faster image generation version.
Z-Image-Base. This is the full base model without any speed-up changes. It's meant for the community to fine-tune or build on for their own uses.
Z-Image-Edit - made for image editing.
Z-Image Turbo hit over 1 million downloads in just a week after its release for ComfyUI workflow.
If you'd like to access this model, you can explore the following possibilities:
LoRA
If you’re generating retro assets, you need this in your workflow.
Workflow
ComfyUI utility nodes for Z-Image model. Features LLM-powered prompt enhancement using the official Z-Image system prompt.
LoRA
AI Toolkit is the ultimate training toolkit for finetuning diffusion models now supports training LoRAs for Z-Image Turbo.
Workflow
An example of a workflow for quantized versions of Z-Image Turbo model.
Weights