devin-lai/flux-3-image-klein-mlx-q4

FLUX.3 Image Klein MLX — Q4

Smallest transformer; larger fidelity tradeoffs. One complete model configuration for local generation and reference-image editing on Apple Silicon. Published by Devin Lai; an independent repacked derivative of Black Forest Labs' distilled checkpoint.

All models and quick start · Original image comparisons · Benchmark report

Original 768 × 512 output, seed 1937. Full caption, settings and comparisons. This image is AI-generated.

What you download

Transformer: q4. Text encoder: q8. The tokenizer, FP32-stored VAE, image-conditioning weights, license files and runnable presets are included. Other variant repositories are not needed. The small inference runtime is installed once from the main repository.

Model component download: 7.40 GiB, plus small metadata and this preview. On the tested MacBook Pro M5 / 32 GiB, q4-16 measured 33.75 s across four workloads and 5.53 GiB maximum process footprint. These are 16-step workload aggregates with three fresh-process repeats per case; download size and runtime memory differ. No minimum supported Mac memory capacity has been established. Methods, per-case ranges and limitations.

Run

Use Apple Silicon and Python 3.12. In a fresh environment:

python3.12 -m venv .flux-env
source .flux-env/bin/activate
python -m pip install "huggingface_hub>=1,<3"
hf download devin-lai/flux-3-image-klein-mlx \
  --include "flux_klein_mlx/*" --include "pyproject.toml" \
  --include "LICENSE-runtime.txt" --include "examples/*" \
  --local-dir flux-klein-runtime
python -m pip install "./flux-klein-runtime[hub]"
flux-klein --download q4 --model-dir ./flux-models/q4 \
  --preset q4-16 --prompt-file ./flux-klein-runtime/examples/teapot.json \
  --width 768 --height 512 --seed 1937 --out ./flux-output/teapot-q4.png

Omit --download for subsequent offline runs. Use --list to see this variant's available presets. Add --image /path/to/reference.png and a full result caption for editing. The runtime can accept up to ten reference images; the local benchmark covers one-reference editing. Captions are used verbatim; no prompt upsampler or content filter is bundled. See the usage guide and follow the upstream license's applicable review/filtering and disclosure terms.

To download the files without installing the runtime, use hf download devin-lai/flux-3-image-klein-mlx-q4 --local-dir ./flux-models/q4. SHA256SUMS covers this repository; component manifests record the weight checksums and MLX quantization settings. config.json identifies the packed layout; model_index.json maps the components used by the runtime. This is not a drop-in Diffusers model or a complete Qwen3-VL checkpoint.

Limitations and provenance

Mixed/Q4 can noticeably change composition, typography and object counts. The Q4 typography example adds a second pencil. They did not pass the stricter full-set fidelity criterion used for the reduced-step quality presets.

This release repacks/quantizes existing weights; it adds no pretraining or fine-tuning dataset. Refer to the upstream model card for training provenance, bias and general limitations. Source revision: 1755314e2216b2a2c5cc0dd899d4e49ee3404d2c. Stored weight precision differs from inference arithmetic. Exact grids, arithmetic, kernels and VAE settings are in presets.json. Conversion and calibration source code are not distributed.

FLUX weights retain the FLUX Kommunity License; text encoder assets retain Apache-2.0. Runtime code is Apache-2.0. Rights to the FLUX model are granted directly by Black Forest Labs under its license. See NOTICE.md for required attribution and the modification statement. This release is not endorsed by Black Forest Labs.

If useful, like the variant and follow devin-lai for updates. Share reproducible results or report runtime problems in the main Community tab.