| --- |
| license: other |
| license_name: flux-1-dev-non-commercial-license |
| license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md |
| language: |
| - en |
| library_name: diffusers |
| base_model: black-forest-labs/FLUX.1-dev |
| tags: |
| - mlx |
| - apple-silicon |
| - flux |
| - quantized |
| - text-to-image |
| --- |
| |
| # FLUX.1-dev — MLX quant matrix (bf16 / Q8 / Q4) |
|
|
| Pre-quantized, **MLX-ready** repackagings of [`black-forest-labs/FLUX.1-dev`](https://huggingface.co/black-forest-labs/FLUX.1-dev) |
| for the [SceneWorks](https://github.com/SceneWorks/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 | Approx. size | |
| |--------|---------|--------------| |
| | Q4 | `q4/` | ~8.7 GB | |
| | Q8 | `q8/` | ~17 GB | |
| | bf16 | `bf16/` | ~31 GB | |
|
|
| All four components are packed in Q4/Q8 — the DiT transformer, the CLIP + T5 text encoders, and the |
| VAE's mid-block attention — using plain asymmetric group-affine quantization (group size 64), |
| byte-identical to the worker's load-time quantization. The bf16 tier is the dense source, mirrored. |
|
|
| ## License & attribution — NON-COMMERCIAL |
|
|
| **These weights are governed by the FLUX.1 [dev] Non-Commercial License v1.1.1** (see `LICENSE.md`), |
| inherited unchanged from the upstream Black Forest Labs release. This repository only re-packages the |
| weights (quantization + MLX layout) and adds no new training. Use is permitted for non-commercial |
| purposes only, per that license. © Black Forest Labs. Original model card: |
| https://huggingface.co/black-forest-labs/FLUX.1-dev |
|
|
