Modelpublic

JoyFusionAI/Qwen-Image-2.1-MLX-4bit

sourceHugging Faceotherupdated 6d agoView on Hugging Face

5likes2.8kdownloads

Model Card

Qwen-Image-2.1 · MLX 4-bit (mflux)

Qwen/Qwen-Image-2.1 quantized to 4-bit for Apple Silicon with mflux (mflux-save -q 4, commit 8c00dab2).

Component	Precision	Size
Transformer (7B single-stream DiT)	4-bit (MLX affine, group 64)	3.8 GiB
Text encoder (Qwen3-VL-8B)	bf16 (mflux keeps it unquantized)	14 GiB
VAE (64-channel)	4-bit linear layers, conv layers bf16	1.2 GiB
Total		~20 GB

Usage

Requires an mflux build that includes Qwen-Image-2.1 support (mflux main at or after 8c00dab2).

mflux-generate-qwen-2.1 \
  --model JoyFusionAI/Qwen-Image-2.1-MLX-4bit \
  --base-model qwen-image-2.1 \
  --prompt "A red fox in a snowy forest holding a wooden sign that says \"Hello\"" \
  --steps 40 --low-ram

--low-ram releases the text encoder after prompt encoding and tiles the VAE decode. Measured on an M1 Max at 768×768: peak memory ~7 GB (vs. 33–36 GB without it), ~6.7 s/step, no visible quality change.

Defaults recommended by the model authors: 40 steps, no guidance. Image sizes must be divisible by 16.

License

These weights are a derivative of Qwen-Image-2.1 and are distributed under the Qwen Research License Agreement. Non-commercial research and evaluation use only; commercial use requires a separate license from the Qwen team. See the original license.