toxicdog/Qwen-Image-2.1-MLX

Qwen Image 2.1 MLX 4-bit

This is a 4-bit MLX conversion of Qwen/Qwen-Image-2.1 for Apple silicon Macs.

Qwen Image 2.1 is a unified image generation and editing model. Its visual generation component has 7 billion parameters and 32 single-stream DiT layers. It supports text-to-image generation, editing with up to 10 reference images, and transparent RGBA output.

The conversion was made primarily for Radiant Canvas, a native macOS app for local image generation and editing. Radiant Canvas uses a custom C++/MLX Qwen Image 2.1 inference engine for fast local inference on Apple silicon. Support for this model is planned in an app update in the next few days.

Quantization

• Framework: MLX
• Quantization: 4-bit affine
• Group size: 64
• Quantized tensors: compatible 2D linear weights
• Preserved tensors: embeddings, normalization parameters, biases, convolution kernels, and other non-linear tensors
• VAE precision: bf16
• Base model: Qwen/Qwen-Image-2.1
• Conversion source: toxicdog/Qwen-Image-2.1 bf16 component files

Each quantized linear layer is stored in MLX format as a packed uint32 weight plus bf16 scales and biases tensors. This is an MLX checkpoint. It is not a drop-in Diffusers or PyTorch checkpoint.

The repository includes the original processor and component configuration files for use by custom runtimes. See quantization_config.json for the machine-readable quantization settings and conversion reports.

Files

Component	File	Precision	Conversion result
Image transformer	diffusion_models/qwen_image_2.1_mlx_q4.safetensors	MLX q4 plus preserved bf16 tensors	200 linear weights quantized, 65 tensors preserved
Qwen3-VL 8B encoder	text_encoders/qwen3vl_8b_mlx_q4.safetensors	MLX q4 plus preserved bf16 tensors	343 linear weights quantized, 407 tensors preserved
VAE	vae/qwen_image_2.1_vae_bf16.safetensors	bf16	Preserved because its weights are convolutional rather than MLX quantized linear layers

The checkpoint was converted with MLX 0.32.2. The JSON files in conversion_reports list every quantized and preserved source tensor.

Qwen Image 2.1

The base model supports:

• text-to-image generation
• image editing with up to 10 reference images
• local edits guided by circles, painted annotations, or masks
• transparent RGBA image generation and editing
• identity preservation for people and products
• multilingual text rendering

For architecture details, examples, and prompt guidance, see the official Qwen Image 2.1 model card and Qwen Image 2.1 repository.

License

This conversion is distributed under the Qwen Research License Agreement, the same license as the base model. The license permits use, modification, and redistribution for non-commercial research or evaluation. Commercial use requires a separate license from Qwen.

The original Qwen attribution and details of the changes made in this conversion are in NOTICE.