mlx-community/Qwen-Image-2512-8bit

Qwen-Image 8-bit MLX

MLX-optimized 8-bit quantized version of Qwen-Image for Apple Silicon.

First MLX port of Qwen-Image! 🎉

Quick Start

pip install mflux

mflux-generate-qwen \
  --model mlx-community/Qwen-Image-2512-8bit \
  --prompt "A photorealistic cat wearing a tiny top hat" \
  --steps 20

Performance

Metric	Value
Size	34GB (8-bit quantized)
Speed	~8.5s/step on M-series Mac
20 steps	~2:50 total

Model Details

• Base Model: Qwen/Qwen-Image (Dec 2025)
• Quantization: 8-bit
• Framework: MLX (Apple Silicon optimized)
• Converted with: mflux 0.14.0

Hardware Requirements

• Apple Silicon Mac (M1/M2/M3/M4/M5)
• ~40GB unified memory recommended for 8-bit

Sample Output

Generated with prompt: "A photorealistic cat wearing a tiny top hat, studio lighting"

All Quantizations

• 8-bit (34GB)
• 6-bit (29GB)
• 5-bit (27GB)
• 4-bit (24GB)
• 3-bit (22GB)

License

Apache 2.0 (same as base model)

Credits

• Original model: Qwen Team
• MLX framework: Apple ML Explore
• mflux: filipstrand