Qwen-Image-2.1 mflux q4
Native MLX/mflux q4 conversion of Qwen/Qwen-Image-2.1 for Apple silicon.
The transformer and Qwen3-VL text encoder are both stored in MLX affine 4-bit
format; the VAE uses the mflux Qwen Image 2.1 layout.
This checkpoint was converted from upstream revision
790c92633540aa0cb11d9abf19eb46d861714758 with mflux 0.20.0 and MLX 0.32.2.
It is intended for the reviewed low-memory loader in Rapid-MLX. Stock mflux
0.20.0 skips text-encoder quantization and cannot load this pack without that
loader change.
Measured during feasibility testing on a Mac Studio:
- 512x512, 40 steps: 4.68 GiB peak MLX memory
- 1024x1024, 4 steps: 5.14 GiB peak MLX memory
- packaged size: approximately 8.9 GB
These measurements used prompt materialization, encoder/transformer eviction, a zero MLX cache, and tiled VAE decoding. Physical 8 GB and 16 GB Mac release qualification is still required.
The original model and this conversion are licensed under Apache-2.0. See
LICENSE and the upstream model.