FLUX.2 [klein] 9B: 4 steps, guidance 1.0, ~29 GB VRAM and non-commercial duties
FLUX.2 [klein] 9B (9B flow + 8B Qwen3 text embedder, 4-step distilled): diffusers Flux2KleinPipeline, bf16, 1024x1024, guidance_scale=1.0, 4 steps, enable_model_cpu_offload. ~29 GB VRAM per card. FLUX Non-Commercial License: filters or manual review required. 4B is Apache-2.0.
FLUX.2 [klein] 9B is Black Forest Labs' larger klein model: a 9B rectified flow transformer with an 8B Qwen3 text embedder, step-distilled to 4 inference steps, handling text-to-image, single-reference editing and multi-reference generation in one model. BFL describes it as sub-second and competitive with much larger models.
Running: the reference implementation is github.com/black-forest-labs/flux2; it also runs in ComfyUI and diffusers. In diffusers (install from source): Flux2KleinPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-9B", torch_dtype=torch.bfloat16), pipe.enable_model_cpu_offload(), then pipe(prompt, height=1024, width=1024, guidance_scale=1.0, num_inference_steps=4, generator=...). The card says it fits in about 29 GB VRAM and is accessible on an RTX 4090 and above; a 24 GB 4090 therefore relies on offloading such as the example's CPU offload.
Licence and duties: the 9B weights are under the FLUX Non-Commercial License (the 4B klein models are Apache-2.0). Filters or manual review must be used with the 9B models; the repository includes NSFW and protected-content filters for inputs and outputs, and the inference code includes an example pixel-layer watermark. The upstream repository requires accepting the terms; this Unsloth mirror is ungated but states the upstream terms are unchanged and still apply. Limitations: no factual information, possibly distorted text, possible bias, prompt following depends on prompting style.
Based on the unsloth/FLUX.2-klein-9B card pinned to revision 455f1273c2f08f1fc93d725c2e6f1d7fc681cd94 and observed on 2026-10-27; the upstream card was not archived in this packet.
Nothing was run here; the 29 GB figure and the speed claims are BFL's, and the licence text was not reviewed beyond the card.
The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.
Structured fields
{
"citations": [
{
"url": "https://huggingface.co/unsloth/FLUX.2-klein-9B/raw/455f1273c2f08f1fc93d725c2e6f1d7fc681cd94/README.md",
"note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 207:517, 1330:1420, 3161:3316, 4749:4843, 7748:10002; exact excerpt SHA256 da287d86c746ff89da8e76a4a9a3d930389acc0911b6c2182b1d2ff652b167d1,65ccb25195344c5f3f8e5ca5e556cfed84cc821d78316d7e4a01af848a66c0d1,43fe373073d5595258be7ff695c52fc5371d1132e02ec50aa4b73b2b779fbf8c,ff4f7f8c"
}
]
}Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | permitted | unknown | Original attributed paraphrase and short functional values only. Source text, images, weights and other assets remain separately governed. |