Alibaba unveiled a new image generation model called Z-Image Turbo with 6B parameters and subsecond-style responsiveness on modern GPUs.

Performance and resource requirements

The model runs efficiently and fits into a 16 GB VRAM footprint, producing a 1024×1024 image in 9 steps that typically takes about 3 seconds.

Image quality and multilingual support

Outputs demonstrate strong realism and fine detail, and iterative prompt refinement often yields the desired composition within a few attempts.

Z-Image Turbo understands Russian and can render text in that language, although very small captions may be rendered inaccurately.

Limitations and behavior

The model shows limited diversity on repeated prompts, sometimes producing a different number of objects than requested and preferring certain facial phenotypes by default.

Many of these tendencies can be mitigated by prompt engineering, including translating prompts to Chinese for improved fidelity in some scenarios.

Availability and tooling

For now, only the Turbo variant has been released as open source, while the undistilled Base and the edit-oriented Edit versions remain pending.

• Comfy has added support and published weights for community integration.
• Workflows and examples for popular UIs are already available to simplify adoption by practitioners.

Developers and artists testing the model report fast runtimes and practical image quality, with further model variants expected in subsequent releases.