Introduction

We already covered the closed FLUX 3 Image stack — boxes, pixel-locked edits, native 4K. Today Black Forest Labs ships the open-weights sibling: FLUX 3 Image Klein, a 5.6B parameter model for text-to-image, instruction edits, and multi-reference generation with up to 10 refs.

Official blog dated October 15, 2026: FLUX 3 Image Klein. Same core controls as the large FLUX 3 Image model — native 4K, pixel-perfect editing, bounding-box layout — sized to run on a single consumer GPU.

What’s in the Drop

Per BFL:

  • Distilled Klein — low-latency generation and editing (the checkpoint behind the BFL API; ~16 steps, no CFG)
  • Klein Base — undistilled checkpoint for fine-tuning, LoRA training, and research (~50 steps, guidance ~3.5)
  • Quantized builds — FP8, INT8, and NVFP4 developed with NVIDIA support
  • License — FLUX Kommunity License (non-commercial + commercial output for qualifying orgs under ~$5M annual revenue; larger commercial self-host needs a commercial license)

BFL’s internal evals put Klein ahead of larger open weights peers (including Qwen-Image 2.1 at 7B) on text-to-image, multi-reference, and editing, while claiming a Pareto win on speed vs quality vs size. Treat that as vendor eval, not an independent arena ranking — but the runnable surfaces are real today.

Hardware Reality Check

From the blog and NVIDIA measurements on the distilled build:

  • Generator weights: ~11.2 GB BF16, ~5.6 GB FP8
  • INT8 path aimed at RTX laptops/desktops with as little as 8 GB VRAM via ComfyUI or Diffusers
  • Rough distilled timings: ~4 s for a 1MP image on RTX 5090 (INT8), under 9 s on RTX 4070 through ComfyUI; ~0.55 s on NVIDIA GB300 (NVFP4)

Peak memory with BFL’s reference code (B200, 1024×1024, upsampler off) is higher than the generator-only numbers — plan headroom for the Qwen3-VL text encoder and optional prompt upsampler.

Where to Run It Today

  • Hugging Face weights: black-forest-labs/flux-3-image-klein (distilled) and flux-3-image-klein-base
  • ComfyUI: native day-0 support per
    — also Comfy-Org / BFL Comfy packages on HF (flux-3-image-klein-comfy)
  • Diffusers: Modular pipeline writeup — Welcome FLUX.3 Image Klein in Diffusers
  • fal: blackforestlabs/flux-3/klein/text-to-image and edit endpoints (planned ~$0.009 at 1K; Base variants also listed)
  • BFL API: flux-3-image-klein with prompt upsampling and pixel restoration — pricing starts at $0.009 per 1K image
  • Training: Ostris AI Toolkit + NVIDIA NeMo AutoModel recipes from day one

Quick Diffusers path (install from git until the release pins catch up):

pip install git+https://github.com/huggingface/diffusers.git

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")

image = pipe(
    prompt="A wide-angle photograph of a lone American bison in a misty geothermal landscape, half hidden by white steam, muted cool palette, soft overcast light.",
    generator=torch.Generator().manual_seed(0),
    output="images",
)[0]
image.save("bison.png")

Prompting Notes Worth Keeping

BFL is blunt: short prompts go generic. Write full scene captions (framing, subject, background, lighting), or use their prompt upsampler (flux-3-image-klein-pup). For layout, append JSON boxes on a 0–1000 grid and tag elements in the caption — boxes guide placement/scale; they are not hard pixel masks.

Edits work best as a complete caption of the target image with the change folded in, not a one-liner like “make it blue.”

How This Differs From What We Already Covered

  • Prior magazine piece FLUX 3 Image: Boxes, 4K & Pixel Edits = the larger / commercial FLUX 3 Image surface (weights “soon” at the time).
  • This drop = open weights Klein (5.6B) you can download, LoRA, and self-host under Kommunity terms — plus the cheap API/fal lane.
  • FLUX.2 Klein articles on the site are the previous generation; don’t mix the model pickers.

Original Source

Also: BFL blog, model page, HF distilled, Diffusers blog, fal text-to-image.

Conclusion

Open weights finally catch the FLUX 3 Image feature set in a 5.6B package — boxes, multi-ref, native 4K, and an 8 GB INT8 story that local creators can actually try tonight. Grab the distilled weights for speed, Base for LoRAs, or hit fal / the BFL API if you want zeros in your VRAM bill.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=flux-3-image-klein-open-weights-5-6b

—Aurelia ♡