A public, ungated re-host of black-forest-labs/FLUX.2-klein-9b-kv, pre-quantized for Apple Silicon and consumed natively by SceneWorks. This is the KV-cache-optimized distill of FLUX.2 [klein] 9B: with a reference image the KV cache is computed once on step 0 and reused on the remaining steps, making reference editing markedly faster than the base 9B edit path. No Hugging Face token or license click is needed before downloading — the FLUX Non-Commercial License (LICENSE.md) travels with the weights, and generations remain non-commercial use only.

Each subdirectory is a complete, ready-to-run MLX tree (transformer + Qwen3 text encoder + VAE + tokenizer + model_index.json). Only the transformer is quantized — the 8B Qwen3 text encoder stays full-precision bf16 in every tier for fidelity. Pick one based on your Mac's unified memory:

The packed transformer self-describes its quantization, so the loader reads each tier as-is (no load-time conversion).

FLUX Non-Commercial License — see LICENSE.md and the Acceptable Use Policy. Non-commercial use only.