FLUX 3 Image Klein is out with open weights: will it run on your GPU, and can you use it commercially?

Back in July, Black Forest Labs announced FLUX 3 but shipped only gated video access, with the image model and open weights still “coming”. On October 15, 2026 the first open piece arrived: FLUX 3 Image Klein, a 5.6B-parameter open-weights model that does text-to-image, image editing and multi-reference generation with up to 10 reference images. For local builders, the questions are simple. Will it fit on my card, how do I load it, and am I allowed to ship with it?

What shipped

Item	Detail (Black Forest Labs / Hugging Face)
Size	5.6B parameters for the image transformer. The Qwen3-VL-4B-Instruct text encoder is separate
Checkpoints	flux-3-image-klein (step-distilled, 16 steps, no CFG) and flux-3-image-klein-base (50 steps with CFG, for fine-tuning)
Builds	BF16, FP8, INT8 and NVFP4, developed with NVIDIA
Features	Text-to-image, editing, up to 10 reference images, bounding-box layout control
Access	Gated on Hugging Face (you agree to share contact info), plus the BFL API from $0.009 per 1K image
Day-one runtimes	ComfyUI, Diffusers, TensorRT-LLM, SGLang Diffusion, vLLM-Omni

How much VRAM you need

BFL says the generator weights take 11.2 GB in BF16 and 5.6 GB in FP8. With the INT8 build it runs on GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers. NVIDIA’s measurements put a 1MP image from the distilled INT8 model at about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 in ComfyUI.

The full pipeline needs more than the weights, because the text encoder and autoencoder load too. Hugging Face’s Diffusers launch post measured peak allocation after inference at:

Build (GPU)	Peak after load	Peak after inference
NVFP4 (RTX 6000)	13.4 GiB	15.3 GiB
FP8 rowwise (H100)	14.35 GiB	16.27 GiB
BF16 (H100)	18.80 GiB	20.72 GiB

The model card reports higher numbers for BFL’s own reference code on a B200 at 1024×1024: 31 GB in BF16, 26 GB in FP8 and 25 GB in NVFP4. The optional prompt upsampler adds about 28 GB. In short, the low-VRAM path is ComfyUI or Diffusers with a quantized build, not the reference repo.

Hardware also decides which quantization actually speeds things up. In Diffusers, fp8r gets fast kernels on compute capability 8.9+ (Ada, Hopper, Blackwell). nvfp4 is fast only on Blackwell with matching flashinfer-python and flashinfer-cubin versions. On other GPUs, the weights are dequantized each step: memory drops, but you get no speedup.

Load it in Diffusers

Hugging Face’s minimal example uses the new modular pipeline:

import torch
from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained("black-forest-labs/flux-3-image-klein")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")
pipe(prompt="a red fox", output="images")[0].save("fox.png")

Install with pip install -U diffusers transformers. You need transformers for the Qwen3-VL text encoder. If memory is tight, the post suggests a quantized transformer via from_single_file, running the text encoder once and freeing it, or ComponentsManager with enable_auto_cpu_offload. On one H100, the distilled model took 1.93 s for a 1024×1024 text-to-image run. A reference image roughly doubles that to 3.87 s, because each 1024×1024 reference adds 4,096 tokens to every transformer call.

Layout control is plain text: append a JSON list of regions after the caption. Boxes are [y_min, x_min, y_max, x_max] on a 0–1000 scale, y first. The model card warns that boxes guide placement and scale but are not hard masks.

The license question

The weights use the FLUX Kommunity License v1.0. It allows non-commercial and research use. BFL’s blog says it also allows free commercial use for individuals and companies with annual revenue under $5M. The model card frames this as commercial output generation by “Qualifying Users” under US$5M gross annualized revenue, including affiliates. It also notes that this exception is not blanket permission for commercial fine-tuning or distribution. Larger companies need BFL’s self-hosted license or the API.

Should you switch?

If you want a small local image model, Klein is now the obvious one to test against ternary options like Bonsai Image 4B. Our Bonsai local setup guide covers the same workflow on smaller hardware. BFL’s comparisons (higher Elo than FLUX.2 [klein] 9B with 38% fewer parameters) come from its own internal evaluations, so run your own prompts before you migrate a production pipeline.