rskulles/z-image-turbo-mflux-q8

Z-Image-Turbo, mflux 8-bit

Z-Image-Turbo by Tongyi Lab (Alibaba), converted to mflux's saved format with 8-bit weights, for image generation on Apple Silicon through MLX. About 10 GB instead of the 20 GB of the original bf16 weights, with no visible loss in the pictures.

This is the copy Crayon Cloud downloads on first run, so its users get a 10 GB download and no quantisation step. Nothing was trained or changed beyond the quantisation.

How it was made

Loaded with mflux 0.19.1 (ZImage(model_config=ModelConfig.z_image_turbo(), quantize=8)) from Tongyi-MAI/Z-Image-Turbo at revision f332072aa78be7aecdf3ee76d5c247082da564a6, then written out with mflux's ModelSaver (save_model(model, 8, path, ZImageWeightDefinition)). The folder holds the transformer (6.1 GB), the Qwen3 text encoder (4.0 GB), the VAE and the tokenizer.

Use

With mflux:

mflux-generate-z-image-turbo --model-path rskulles/z-image-turbo-mflux-q8 --prompt "a lighthouse at dusk" --steps 8 --seed 7

In Python:

from mflux.models.common.config.model_config import ModelConfig
from mflux.models.z_image import ZImage

model = ZImage(model_config=ModelConfig.z_image_turbo(), model_path="rskulles/z-image-turbo-mflux-q8")
image = model.generate_image(seed=7, prompt="a lighthouse at dusk", num_inference_steps=8, width=1024, height=1024)

Z-Image-Turbo is guidance-distilled: 8 or 9 steps, no guidance, no negative prompt.

Licence

Apache 2.0, the same as the original weights. Copyright Tongyi Lab; conversion by Roy Skullestad. If mflux changes its saved format, this repository will be refreshed; the original repository is always the source of truth.