qwen-image-edit-4bit

This repository contains MLX-Gen saved weights for Qwen/Qwen-Image-Edit. The checkpoint is designed for local Apple Silicon inference with mlx-gen.

It uses the mflux/MLX saved-weight layout and MLX quantization tensors. It is not a Diffusers or Transformers from_pretrained() checkpoint.

Source Model

Original model: Qwen/Qwen-Image-Edit.

License and Access

This quantized derivative follows the Apache 2.0 license of the source model.

Quantization

This is a mixed q4/q8 checkpoint for Qwen Image generation and editing. Fully q4 Qwen checkpoints can lose coherent generative behavior, so MLX-Gen uses a mixed policy:

• q4 for most Qwen transformer attention, feed-forward, and projection linears.
• q8 for Qwen *.img_mod_linear transformer modulation layers.
• q4 for group64-compatible Qwen text-encoder language linears.
• q8 for group64-compatible Qwen text-encoder visual linears.
• BF16 for the VAE, norms, embeddings, and linears that are not MLX group64-compatible.

See the MLX-Gen quantization docs for the current mixed q4/q8 policy and compatibility notes.

Compatibility

Requires mlx-gen >= 0.18.2.

Generated with mlx-gen 0.18.2.

Use the mlxgen command and Python import path for new MLX-Gen projects.

Usage

python -m pip install -U mlx-gen

mlxgen download --model AbstractFramework/qwen-image-edit-4bit

mlxgen generate \
  --model AbstractFramework/qwen-image-edit-4bit \
  --image path_to_image.png \
  --prompt "Your edit prompt here" \
  --steps 20 \
  --seed 42 \
  --output edited.png

Attribution

MLX-Gen is based on mflux by Filip Strand and the original mflux contributors. This model card is generated by MLX-Gen so derived checkpoints keep that attribution visible.

Quantized and contributed by @lpalbou.