Qwen-Image-mlx-4bit
4-bit MLX weights for Qwen-Image diffusion pipeline on Apple Silicon, converted using mflux.
Converted by: SirSahOl
Base Architecture: Qwen/Qwen-Image-2512 (60-block DiT topology)
Framework: mflux
Quantization: 4-bit
Format: .safetensors
License: Tongyi Qianwen License
Model Details
| Attribute | Value |
|---|---|
| Pipeline Type | Diffusion Transformer (Text-to-Image) |
| DiT Backbone | 60 Transformer Blocks (~7.0B Parameters) |
| Composite Pipeline | DiT Transformer + Text Encoder + VAE (~8.2B Total Parameters) |
| Quantization | 4-bit |
| Format | Apple Silicon MLX (.safetensors) |
| Active VRAM Footprint | ~33.5 GB |
Quick Start (Apple Silicon)
1. Installation
pip install -U mflux
2. CLI Generation
mflux-generate-qwen \
--model SirSahOl/Qwen-Image-mlx-4bit \
--prompt "A cinematic portrait of a robotic artisan crafting glass in a neon workshop, 8k, photorealistic" \
--steps 25 \
--seed 42
3. Python API
from mflux import QwenImagePipeline
pipeline = QwenImagePipeline.from_pretrained("SirSahOl/Qwen-Image-mlx-4bit")
image = pipeline.generate(
prompt="A serene mountain lake at sunrise with golden mist, ultra realistic",
num_inference_steps=25,
)
image.save("output.png")
Multi-Quantization Variants
| Precision | Repository | Key Advantage |
|---|---|---|
| 4-bit | SirSahOl/Qwen-Image-mlx-4bit | Fastest generation and lowest VRAM requirement (~33.5 GB). |
| 8-bit | SirSahOl/Qwen-Image-mlx-8bit | Near-lossless visual quality and prompt adherence (~49.7 GB). |
| 16-bit | SirSahOl/Qwen-Image-mlx-16bit | Full unquantized bfloat16 reference weights (~72.6 GB). |
Architectural Note
These weights represent the 60-block transformer topology (Qwen/Qwen-Image-2512 base) supported by mflux. Native 32-layer Qwen-Image-2.1 weights will be released once upstream mflux merges support for the newly announced 32-layer architecture.