| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen-Image-2.1 |
| library_name: mflux |
| tags: |
| - mlx |
| - mflux |
| - text-to-image |
| - image-to-image |
| - qwen-image |
| --- |
| |
| # 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](https://huggingface.co/Qwen/Qwen-Image-2.1). |
|
|
