otherimage-genRuns locally5.6B (image transformer)

FLUX 3 Image Klein: Specs, License & How to Run

Black Forest Labs 5.6B open image transformer; text-to-image, editing, up to 10 refs, native 4K.

Official Black Forest Labs FLUX 3 Image Klein open-weights launch collage — Official project image. Source · License: Black Forest Labs product image.

License: other
Parameters: 5.6B (image transformer)
HF likes: 238 (2026-10-16)
First seen: 2026-10-13

TL;DR. FLUX 3 Image Klein is a image-gen model from Black Forest Labs. Parameters: 5.6B (image transformer).License: other. It is listed under Midjourney.

Developer	Black Forest Labs
Type	image-gen
Architecture	5.6B image transformer; separate Qwen3-VL-4B-Instruct text encoder (Apache-2.0). Distilled 16-step (no CFG) + Base 50-step (CFG 3.5)
Parameters	5.6B (image transformer)
Modalities	image, text
License	other — FLUX Kommunity License v.1.0 (card license_name flux-kommunity-license). Non-commercial use, plus commercial output generation for Qualifying Users with less than US$5,000,000 gross annualized revenue including affiliates. Text encoder Qwen3-VL-4B-Instruct is Apache-2.0. Not legal advice.
Hardware	RTX 3090 (BFL docs) — BFL docs (checked 2026-10-16) say Klein fits consumer GPUs down to an RTX 3090, via ComfyUI or Diffusers. The BFL blog says INT8 runs on GeForce RTX GPUs with as little as 8 GB VRAM; generator weights 11.2 GB BF16 / 5.6 GB FP8. The HF card quotes peak memory on a B200 at 1024×1024 with the prompt upsampler off — 31 GB BF16, 26 GB FP8, 25 GB NVFP4 — and says the upsampler adds about 28 GB. Those are vendor figures. forkalt has not measured VRAM. (not independently tested)
Hugging Face likes	238 (as of 2026-10-16)
Model card	https://huggingface.co/black-forest-labs/flux-3-image-klein
First seen	2026-10-13

FLUX 3 Image Klein is an open-weight image model from Black Forest Labs. Official hosts we opened on 2026-10-16: the distilled card black-forest-labs/flux-3-image-klein, the Base card black-forest-labs/flux-3-image-klein-base, the prompt upsampler black-forest-labs/flux-3-image-klein-pup, the FLUX Kommunity License, reference code black-forest-labs/flux-3-image-klein, the launch post FLUX 3 Image Klein (15 October 2026), docs.bfl.ml, and the Diffusers post Welcome FLUX.3 Image Klein.

This identification page hangs off open Midjourney-class alternatives. That mapping is editorial. We have no sourced bake-off against Midjourney or Ideogram.

The BFL blog (15 October 2026) calls it a 5.6B open-weights model for text-to-image, image editing, and multi-reference generation with up to 10 reference images, plus native 4K, instruction editing, and bounding-box layout control. Hugging Face pipeline_tag is image-to-image. Those are copied labels, not a ranking.

What we verified (2026-10-16)

Field	Value	Source
Distilled HF likes / downloads	238 / 3,241	HF API `black-forest-labs/flux-3-image-klein`
Card created / last modified	2026-10-13T10:58:40Z / 2026-10-15T22:57:37Z	same
Gated	auto	HF API
License (card)	other / `flux-kommunity-license`	HF `cardData`
LICENSE file	FLUX Kommunity License v.1.0	LICENSE.md
Pipeline tag	image-to-image	HF API
Image-transformer params	5.6B (text encoder separate)	Blog + HF README
HF safetensors (distilled)	5,548,681,984 BF16 (~5.55B)	HF `safetensors.parameters.BF16`
Text encoder	Qwen3-VL-4B-Instruct (Apache-2.0)	HF README
Distilled	16 steps, no CFG	HF README + docs
Base	50 steps, guidance 3.5	same
Capabilities	text-to-image, instruction editing, ≤10 refs, bounding boxes, native 4K	Blog + docs + card

Distilled, Base, and the prompt upsampler

Checkpoint	Steps	Guidance	Built for
flux-3-image-klein	16	none	generation / editing; BFL API
flux-3-image-klein-base	50	3.5	fine-tune / LoRA / research

The HF README and BFL docs say do not send a short prompt as typed: Klein is trained on full scene captions. The first-party upsampler is black-forest-labs/flux-3-image-klein-pup. The reference CLI loads it by default. Diffusers does not call it on its own. The card says the upsampler adds about 28 GB on an 80 GB GPU; that is a vendor figure, not a forkalt measurement.

How to run (from official docs, not yet tested)

BFL docs name ComfyUI and Diffusers. Day-one names on the blog also include TensorRT-LLM, SGLang Diffusion, and vLLM-Omni. A Diffusers sketch copied from the docs / HF blog (checked 2026-10-16):

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

from diffusers import ModularPipeline
pipe = ModularPipeline.from_pretrained("black-forest-labs/flux-3-image-klein")

For LoRA / fine-tune, the same pages point at black-forest-labs/flux-3-image-klein-base. forkalt has not run these commands.

Hardware (vendor)

Docs: consumer GPUs down to an RTX 3090. Blog (vendor): INT8 from about 8 GB VRAM; generator weights 11.2 GB BF16 / 5.6 GB FP8. HF card peak on a B200, 1024×1024, upsampler off: 31 / 26 / 25 GB (BF16 / FP8 / NVFP4). Not independently tested.

License

LICENSE.md is FLUX Kommunity License v.1.0. Hugging Face cardData.license is other. We do not invent a more specific SPDX id.

Paraphrase only, not legal advice: non-commercial use, plus commercial output generation by Qualifying Users (you plus affiliates under US$5,000,000 gross annualized revenue), subject to the license, including content-filter / disclosure conditions in §2.e. Other commercial uses need a separate BFL license (bfl.ai/licensing). The Qwen3-VL-4B-Instruct text encoder is Apache-2.0. The card warns that the commercial-use exception is not blanket permission for commercial fine-tuning or distribution.

What we did not verify

BFL blog Elo deltas (for example +91 Elo text-to-image versus FLUX.2 [klein] 9B) and “ranks first / above Ideogram V4.5” sentences are vendor/self-reported. We do not treat them as a Midjourney or Ideogram verdict. Latency charts on the blog use mixed timing methods (the blog says so). We have not measured VRAM or image quality.

Related image weights on this site: Qwen-Image-2.1-Turbo and Iris-3B on Midjourney-class alternatives.

Deploy options

local: Hugging Face weights (distilled)
local: Hugging Face weights (Base)
local: GitHub reference code
hosted: Black Forest Labs (API / playground, named on the blog)

Comparisons

FLUX 3 Image Klein vs Qwen-Image-2.1-Turbo

FAQ

Are FLUX 3 Image Klein weights out?

Yes. Hugging Face black-forest-labs/flux-3-image-klein is the distilled checkpoint (gated: auto; created 2026-10-13; 238 likes and 3,241 downloads in the 2026-10-16 API check; pipeline_tag image-to-image; card license “other” / license_name flux-kommunity-license). The Base checkpoint for fine-tuning and LoRA is black-forest-labs/flux-3-image-klein-base. Official code is github.com/black-forest-labs/flux-3-image-klein. The BFL blog is dated 15 October 2026.

What is Distilled versus Base?

Same 5.6B image transformer. Distilled is the 16-step checkpoint (no CFG) for generation and editing, and is the checkpoint behind the BFL API. Base is undistilled, 50 steps, guidance 3.5, for fine-tuning, LoRA, and research. The 5.6B label is the image transformer; the text encoder is separate (HF card: Qwen3-VL-4B-Instruct).

Can I use it commercially?

LICENSE.md is FLUX Kommunity License v.1.0. The card and license allow non-commercial use, plus commercial output generation by Qualifying Users with less than US$5,000,000 gross annualized revenue including affiliates, subject to the license terms. The Qwen3-VL-4B-Instruct text encoder is Apache-2.0. Not legal advice; read LICENSE.md.

Does it match Midjourney or Ideogram quality?

We do not claim that. This page identifies an open-weight image checkpoint. BFL blog Elo and “ranks first” sentences versus named open and closed models (including Ideogram V4.5) are vendor/self-reported. There is no sourced bake-off against Midjourney on this page.

Sources & verification

Last verified 2026-10-16 · Last updated 2026-10-16· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.