Qwen-Image-2.1-MLX-4bit — for MindCraft Studio
This repository contains the 4-bit quantized version of Qwen-Image 2.1, optimized for Apple Silicon (macOS) using MLX.
Qwen-Image 2.1 is an advanced visual foundation model featuring strong multilingual prompt understanding, native transparent RGBA image generation, and instruction-based multi-image editing (supporting up to 10 reference images).
- Original Model: Qwen/Qwen-Image-2.1
- Used by: MindCraft Studio — A local-first AI image generation & canvas powerhouse for macOS
- Framework: Apple MLX
- Quantization: 4-bit weight quantization (~11.6 GB)
- License: Qwen Research License Agreement (same as upstream)
Features
- Apple Silicon Native: Optimized specifically for Unified Memory architectures (M1/M2/M3/M4/M5) via MLX.
- Instruction-Based Image Editing: Edit existing images using natural language instructions with support for 1–10 input reference images.
- Native RGBA Transparency Support: Generate cutout stickers and transparent PNGs natively with prompt triggers (e.g.,
"This is an RGBA image with transparency... The image has alpha channel and the background is transparent."). - Bilingual Prompt Understanding: Excellent prompt adherence in both English and Chinese.
How to Use
In MindCraft Studio (Recommended)
No manual conversion or shell scripting required:
- Download and open MindCraft Studio.
- Go to Settings → Model Library.
- Select Qwen-Image 2.1 (4-bit) and click Download / Install.
- The model will download directly from this repository and become available in both the generation sidebar and the endless canvas node engine.
License
This model follows the original upstream license: Qwen Research License Agreement.
Links & Credits
- Original Base Model: Qwen/Qwen-Image-2.1
- Qwen GitHub: https://github.com/QwenLM/Qwen-Image-2.1
- MindCraft Studio: https://themindstudio.cc/mindcraft