Qwen-Image (2512) — MLX quant matrix (bf16 / Q8 / Q4)
Pre-quantized, MLX-ready repackagings of Qwen/Qwen-Image-2512
for the SceneWorks native Apple-Silicon worker (mlx-gen).
Each tier is a complete, self-contained turnkey snapshot that loads directly with no in-app
conversion peak (epic 8506).
| Tier | Subdir | What's packed | Approx. size |
|---|---|---|---|
| Q4 | q4/ |
transformer packed to 4-bit (group 64); dense bf16 TE + VAE | ~26 GB |
| Q8 | q8/ |
transformer packed to 8-bit (group 64); dense bf16 TE + VAE | ~36 GB |
| bf16 | bf16/ |
dense bf16 transformer + TE + VAE (full precision) | ~55 GB |
Only the transformer is quantized. The Qwen2.5-VL text encoder is skip_quantization
(quantization degrades its semantics) and the VAE is all-convolution (no quantizable leaves), so both
stay dense bf16 in every tier. Q4/Q8 use plain asymmetric group-affine quantization (group size 64),
byte-identical to the worker's load-time quantization.
License & attribution
Apache License 2.0, inherited from the upstream Qwen-Image-2512 release. The model weights are
© Alibaba Cloud / the Qwen team; this repository only re-packages them (quantization + MLX layout) and
adds no new training. See LICENSE. Original model card: https://huggingface.co/Qwen/Qwen-Image-2512