What Is FLUX 3 Image Klein? 5.6B Open-Weights Image Model

What Is FLUX 3 Image Klein? 5.6B Open Weights, Native 4K, $0.009 per 1K Image

FLUX 3 Image Klein is a 5.6-billion-parameter open-weights image generation and editing model that Black Forest Labs released on October 15, 2026. A single set of weights handles text-to-image, single-image editing, and multi-reference generation with up to 10 reference images. It supports native 4K output and bounding-box layout control. The weights are on Hugging Face under the FLUX Kommunity License, and the BFL API serves the distilled version at $0.008 to $0.036 per image. With the INT8 build, it runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM.

Key Takeaways

• 5.6B parameters, generator only. The count covers the image transformer. The Qwen3-VL-4B-Instruct text encoder is separate.
• Two checkpoints. Klein Base (undistilled, 50 steps, CFG 3.5) is for fine-tuning. Klein (Distilled) runs 16 steps without guidance and is the checkpoint behind the BFL API.
• Native 4K and 10 references. Output resolution ranges from 768sq to 4K, with up to 10 reference images per request.
• API price from $0.008. A 1K image costs $0.009 and a 4K image $0.036. The full FLUX 3 Image model charges $0.048 and $0.607 for the same sizes.
• Commercial use has a revenue limit. The FLUX Kommunity License allows free commercial use only below $5 million in annual revenue, per VentureBeat.
• Early community results are mixed. Vendor evaluations rank it first among open models, but early hands-on reports flag weak pose and viewpoint edits.

What Is FLUX 3 Image Klein?

FLUX 3 Image Klein is the compact, open-weights member of Black Forest Labs' FLUX 3 family. BFL introduced FLUX 3 on July 23, 2026 as a multimodal foundation model trained jointly on images, video, and audio (see the FLUX 3 announcement and our FLUX 3 omni-model overview). The full FLUX 3 Image model launched on the API on October 1, 2026. Klein followed two weeks later.

BFL's release notes describe Klein as "a 5.6B parameter version of FLUX 3 Image with open weights" that supports the same features. The problem it addresses is deployment. Teams with data-sovereignty, on-device, or fine-tuning needs can run it on their own hardware instead of sending every request to a hosted service.

The model is fully released. The official launch post went out on October 15, 2026. A community observer spotted the model live on the BFL API minutes before the weights appeared.

Source: @cheatyyyy

FLUX 3 Image Klein at a Glance

Attribute	FLUX 3 Image Klein
Developer	Black Forest Labs (Germany)
Type	Open-weights image generation and editing model (transformer)
Release date	October 15, 2026
Parameters	5.6B image transformer; full pipeline count undisclosed
Text encoder	Qwen3-VL-4B-Instruct (separate)
Inputs	Text prompt plus up to 10 reference images
Output	One image, 768sq to 4K
Context window	Not yet confirmed (prompt length limit not published)
Checkpoints	Klein Base (50 steps, CFG 3.5); Klein (Distilled) (16 steps, no CFG)
Quantized builds	FP8, INT8, NVFP4 (developed with NVIDIA)
Weight size	11.2 GB BF16; 5.6 GB FP8 (generator only)
Minimum VRAM	8GB on GeForce RTX with INT8 build
API pricing	$0.008 (768sq) / $0.009 (1K) / $0.011 (1.5K) / $0.014 (2K) / $0.036 (4K)
Availability	BFL API, Hugging Face, ComfyUI, Diffusers, NVIDIA NIM
License	FLUX Kommunity License; separate commercial license for larger companies

How FLUX 3 Image Klein Works

Five design choices define how Klein operates.

Shared multimodal learning. Klein draws on FLUX 3's training across images, video, and audio. According to the Klein launch blog, images contribute spatial detail, video adds motion and physical dynamics, and audio adds information about events. BFL says this grounding makes outputs "feel grounded in the real world." BFL trained Klein on NVIDIA GB200 GPUs in an NVL72 system.

Klein Base vs Klein (Distilled). Klein Base is the undistilled checkpoint for LoRA training and research. Klein (Distilled) is step-distilled for low latency. A LoRA trained on Klein Base also loads into Klein (Distilled), according to BFL's release notes.

The Prompt Upsampler is mandatory. The Klein model card says plainly: "Do not use the model without a prompt upsampler." Klein expects full scene captions, and short prompts produce generic output. BFL ships a 9B prompt upsampler based on Qwen3.5. Any LLM given BFL's system prompt can do the same job.

Bounding-Box Layout Control. Users append a JSON list of boxes to the prompt. Each box uses [top, left, bottom, right] coordinates on a 0 to 1000 grid. Tagged elements can be placed, recolored, moved, replaced, or removed. Boxes guide placement and scale but do not act as a hard pixel mask, as the NVIDIA NIM model card notes.

API compatibility with FLUX 3 Image. A Klein request takes the same fields as a FLUX 3 Image request: prompt, images, aspect_ratio, resolution, and the box format. Moving a request to Klein only requires switching the endpoint.

One counting caveat matters. A community breakdown adds the 4B text encoder to the 5.6B generator and arrives at roughly 9.6B parameters for inference. BFL states that its comparisons count generator parameters "as each publisher reports them."

What You Can Do With FLUX 3 Image Klein

BFL's model page highlights three output categories: photography (skin, fabric, weather, light), illustration (woodblock, mosaic, clay, pixel art, anime), and type and design. Its demo layout builds a postage stamp from five boxed elements, including a cat, a crescent moon, and kanji text.

Showcased edits include:

• Turning a summer street into winter after light snowfall.
• Freezing a waterfall into ice with icicles.
• Tying a red bandana on a dog without altering anything else.
• Placing a convertible on an empty desert road.

Practical workloads follow from that. VentureBeat lists product photography, repeatable brand assets, and internal design tools as candidates for evaluation, while noting these are not reported customer deployments. BFL also points to on-device editing on laptops and to teams that need to keep data on their own infrastructure.

Fine-tuning is supported from day one. Diffusers and Ostris's AI Toolkit (arch: flux3) support LoRA training, and NVIDIA's NeMo AutoModel ships training recipes.

According to early community testing, Klein handles text rendering and simple object edits well. It struggles with pose changes, viewpoint shifts, and compositing multiple references.

How FLUX 3 Image Klein Compares

Model	Params (generator)	Open weights	API price, 1K	Notes
FLUX 3 Image Klein	5.6B	Yes	$0.009	Native 4K, 10 refs, boxes
FLUX 3 Image	Not yet confirmed	Not yet confirmed	$0.048	$0.607 at 4K
FLUX.2 [klein] 9B	9B	Yes	$0.015 (first MP)	Klein is +91 / +70 Elo
Qwen-Image 2.1	7B	Yes	Not yet confirmed	Ranked below Klein by BFL
Krea 2	12.9B	Not yet confirmed	Not yet confirmed	No image editing

In BFL's internal evaluations, Klein ranks first in text-to-image, multi-reference generation, and image editing among open models. It also rates above two closed models, Ideogram V4.5 and FLUX.2 [pro]. FLUX 3 Image Klein is the smallest model in BFL's own comparison at 5.6B parameters, against 7B for Qwen-Image 2.1 and 12.9B for Krea 2. These are vendor self-reported results.

Independent data is limited so far. According to the same community summary, third-party platform Voyager ran 200 text-to-image and editing tests, blind-judged by Claude, Gemini, and GPT. Klein scored 6.93 against 7.20 for Nano Banana 2.1 (covered in our Nano Banana 2.1 overview), at about one fifth of the per-image cost. Some Reddit users reported that pose edits performed worse than in the previous 4B model. Those reports are unverified, and some testers may have skipped the required Prompt Upsampler.

Klein's vendor numbers are strong. Early community testing suggests it does better at text and simple object edits than at pose changes or multi-image compositing.

Availability: How to Access FLUX 3 Image Klein

Klein is available through five main channels:

• BFL API. The distilled model, with prompt enhancement and native 4K. VentureBeat notes that the hosted pipeline includes BFL's prompt processing, so API results may differ from a bare local checkpoint.
• Hugging Face weights. The distilled flux-3-image-klein repo and the Base repo are public, but users must accept access conditions before downloading.
• ComfyUI. A repackaged ComfyUI repo holds BF16, INT8, and NVFP4 transformers, the text encoder, the INT8 upsampler, and the autoencoder.
• Diffusers. It ships as a Modular Diffusers pipeline. Per the Hugging Face launch blog, the distilled checkpoint produces a 1024 × 1024 image in about 2 seconds on an H100.
• Serving stacks. SGLang Diffusion, vLLM-Omni, and NVIDIA NIM had day-0 support.

Source: @cheatyyyy

Companies above the revenue threshold that want to self-host should contact BFL sales about a commercial license. Readers who mainly need hosted access to the larger family can compare the full Flux 3 model page. That page covers FLUX 3, not Klein.

What We Don't Know Yet

• Full pipeline size. NVIDIA's card says the total parameter count and the transformer configuration are undisclosed.
• 8GB performance. BFL's speed charts do not test an 8GB GPU or 4K output, as VentureBeat points out.
• Exact license text for Klein. VentureBeat's analysis relied partly on the published Kommunity license for FLUX 3 Action. Full Klein-specific terms need confirming.
• Independent benchmarks. Only BFL's internal Elo results and one third-party test exist so far.
• Prompt length limits. These have not been published.

Frequently Asked Questions

Is FLUX 3 Image Klein open source?

No, it is open weights rather than open source. Both checkpoints are on Hugging Face under the FLUX Kommunity License. That license permits non-commercial use and includes a commercial-output exception for Qualifying Users, subject to its terms.

How much does FLUX 3 Image Klein cost?

The BFL API charges $0.008 to $0.036 per image, depending on resolution. A 1K image costs $0.009, a 2K image $0.014, and a 4K image $0.036. Text-to-image and editing are priced the same.

What GPU do I need to run FLUX 3 Image Klein?

An NVIDIA GeForce RTX GPU with 8GB of VRAM can run it using the INT8 build. The generator weights take 11.2 GB in BF16 and 5.6 GB in FP8. BFL reports about 4 seconds per 1MP image on an RTX 5090.

Can I use FLUX 3 Image Klein commercially?

Yes, if your annual revenue is below $5 million, according to VentureBeat. Larger organizations need a separate license to self-host, or they can use the BFL API.

How many parameters does FLUX 3 Image Klein have?

The image transformer has 5.6 billion parameters. That figure leaves out the Qwen3-VL-4B-Instruct text encoder and the autoencoder. The full pipeline count is undisclosed.

What is the difference between FLUX 3 Image Klein and FLUX.2 Klein?

In BFL's evaluations, FLUX 3 Image Klein gains 91 Elo in text-to-image and 70 Elo in editing over FLUX.2 [klein] 9B, with 38% fewer parameters. It also adds native 4K output, bounding boxes, and 10-image references.

Is FLUX 3 Image Klein better than Nano Banana 2.1?

Not in one early test. Voyager's 200-prompt blind evaluation, as summarized by a community post, scored Klein at 6.93 and Nano Banana 2.1 at 7.20. Klein cost about one fifth as much per image. The result has not been independently verified.

What to Watch Next

Three signals will shape this page's next update. The first is independent benchmarks that use the Prompt Upsampler correctly, which should settle the gap between vendor and community results. The second is measured performance on 8GB GPUs at 2K and 4K. The third is the full Klein license text and how BFL defines Qualifying Users.

Try related models on Kie.ai: Qwen Image 2.1, Nano Banana 2.1 and FLUX.2.