ZImage MLX Q4 conversion of Qwen-Image-2.1-Turbo
This is a 4-bit MLX conversion of Qwen/Qwen-Image-2.1-Turbo for the zimage.swift Qwen Image 2.1 pipeline. Built with Qwen. The original Qwen Research License applies; see LICENSE and NOTICE. Use is limited to non-commercial research and evaluation unless you obtain a separate commercial license from Qwen.
The text encoder and transformer use affine 4-bit quantization with group size 64. The VAE is stored in MLX layout as float32. The checkpoint's eight sampling sigmas are retained in model_index.json and used automatically when no explicit schedule is supplied. Component SHA256 digests and source paths are recorded in q4_components.json.
The Diffusers-style package stores model.safetensors in each component directory. The ComfyUI-style package stores the same converted weights as three independently selectable safetensors files. Both formats require the included processor and scheduler files when used with zimage.swift.
Example with the Diffusers-style package:
ZImageCLI qwen-image-2.1 --model /path/to/package \
--prompt "a red apple on a table" --width 256 --height 256 \
--steps 8 --seed 42 --output apple.png
For the component package, pass its root as --model and the three files via --text-encoder, --transformer, and --vae.
Both formats were tested locally on an Apple M2 with 16 GB unified memory at 256 x 256, seed 42, eight steps. They produced byte-identical PNGs. The source model card demonstrates 1680 x 2512 text-to-image and 2048 x 2048 image editing; those resolutions have not yet been validated on this 16 GB machine.
Quantized