You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a Core AI conversion of FLUX.2 [klein] 9B by Black Forest Labs, under the FLUX Non-Commercial License (LICENSE.md): non-commercial, non-production use only. You must use content filters or review output before distributing it, and disclose AI generation where the law requires. By agreeing you accept that licence.
Log in or Sign Up to review the conditions and access this model content.
FLUX.2 [klein] 9B — Core AI (float16)
Black Forest Labs' FLUX.2 [klein] 9B converted to a Core AI diffusion bundle, 1024 × 1024,
4 steps — the "quality" option of the imagine tool. Text-to-image only, square only.
Licence — read this first
- FLUX Non-Commercial License v2.1 (
LICENSE.md): non-commercial and non-production use only. Rights come directly from Black Forest Labs to you, not from us. - §2(e): you must either use content filters to prevent unlawful or infringing output, or
review output before distributing it — and disclose AI generation where the law requires.
imaginewrites the IPTC digital-source-type field (trainedAlgorithmicMedia) into every file. - BFL's Attribution Notice and our modification statement are in
NOTICE. This is not an official Black Forest Labs product. - Components (all in BFL's release, all under this licence): the 9B transformer, the Qwen3-8B text encoder (Apache-2.0 upstream) and the FLUX.2 VAE.
Conversion
Apple's recipe, apple/coreai-models (BSD-3) at 52dbd93,
float16, no quantization, --single-function, 1024: --experimental (9B is not in Apple's list),
one component at a time so the 9B transformer export never held the text encoder (≈14 GB peak).
Files
| size | |
|---|---|
metadata.json, tokenizer/, vae_bn_mean.npy, vae_bn_var.npy |
small |
TextEncoder.aimodel (Qwen3-8B, the layers FLUX uses) |
11 GB |
Transformer.aimodel |
17 GB |
VAEDecoder.aimodel |
95 MB |
Measured (macOS 27, M3 Max, 36 GB, GPU, 1024 × 1024, imagine)
| Measure | Result |
|---|---|
one picture, whole command, parts loaded one at a time (--low-memory) |
99–137 s, median ≈110 s |
| peak memory footprint, one at a time | 4.1 GB |
| all parts resident at once | 176 s — slower, memory pressure on 36 GB |
| first run (compiles for this Mac's GPU, cached per program) | 304 s |
| 4B on the same prompts | 44–53 s |
Against 4B on the same 9 prompts and seeds: a tie on photos, products, landscapes and counting; 4B was better on an app icon; 9B was clearly better at longer text ("SOUND EFFECTS PACK" nearly right where 4B wrote gibberish). That is why it is offered, opt-in.
Use
imagine model accept quality # once: shows these terms, records your acceptance
imagine --quality "a coffee shop sign that says FRESH BREAD DAILY" -o sign.png
Downloading needs a Hugging Face account that has accepted this repository's terms and a token
(hf auth login); imagine says so with exit 3 when either is missing.
Model tree for arraypress/imagine-flux2-klein-9b
Base model
black-forest-labs/FLUX.2-klein-9B