Image GenerationOpen ModelsDeveloper Tools

FLUX 3 Image Klein

Through Black Forest Labs' API, FLUX 3 Image Klein costs $0.008 per image at 768x768, $0.009 at 1K, $0.011 at 1.5K, $0.014 at 2K, and $0.036 at 4K, the same for generation and editing. The weights are a free download, but commercial use of outputs is free only for organizations under US$5M in annual revenue (including affiliates); larger companies need a license from BFL. Checked November 7, 2026. Check official pricing →

Quick Answer

FLUX 3 Image Klein is the model to look at when you want FLUX-style image generation and editing without paying top-tier prices per image, or when you want the images made on your own hardware. Black Forest Labs (BFL) released it on October 15, 2026 as the smaller, open-weights member of its FLUX 3 Image family. It generates images from text, edits existing images, combines up to 10 reference images, and accepts bounding boxes to place elements in a layout.

You can use it two ways. BFL’s API charges $0.009 for a 1K image. Or you can download the weights from Hugging Face and run them yourself, as long as the FLUX Kommunity License fits your business: free commercial use of outputs stops at US$5 million in annual revenue.

Model Specifications

FLUX 3 Image Klein
DeveloperBlack Forest Labs
ReleasedOctober 15, 2026
Image model size5.6B parameters (the image transformer; Hugging Face lists it as 6B)
Other componentsQwen3-VL-4B-Instruct text encoder (Apache 2.0), image autoencoder, optional 9B prompt upsampler
CheckpointsKlein (distilled, 16 steps, no guidance; the one behind BFL’s API) and Klein Base (50 steps, guidance 3.5; for fine-tuning and research)
PrecisionsBF16, FP8, NVFP4; an INT8 build made with NVIDIA for ComfyUI
TasksText-to-image, image editing, multi-reference (up to 10 images), bounding-box layout
API resolutions768x768 up to 4K
LicenseFLUX Kommunity License v1.0 (gated download)

The “5.6B” figure covers only the image transformer. The text encoder and the optional prompt upsampler are separate models, and they count toward memory when you run it locally.

What It Does Well

Generation and editing in one model. The same checkpoint creates images and edits them. You can describe a change (“make the jacket red, keep everything else”), and BFL’s API charges the same for an edit as for a new image.

References. You can pass up to 10 reference images, for example a product shot, a logo, and a style example, and ask for a new composition that uses them. That suits product mockups, consistent characters across a series, and putting the same item into different scenes. Consistency is good, not guaranteed; check details such as logos and text in every output.

Layout control. Bounding boxes tell the model roughly where each element should go, which helps with banners, thumbnails, and mockups that need a fixed structure. BFL’s model card is clear that boxes guide placement but are not hard masks, so an edit can still change content outside the box.

Prompts. The model expects full scene descriptions. By default the reference code, Diffusers, and ComfyUI run a separate 9B prompt upsampler that expands a short request into a detailed caption. Short prompts without it tend to give generic results; you can also use another language model with a system prompt to write the captions.

Speed

BFL’s own charts, at 1024x1024 with the 16-step distilled checkpoint, show about 4 seconds to generate and 7 seconds to edit on an RTX 5090 with the INT8 build, and about 9 seconds to generate and 19 to edit on an RTX 4070. Through BFL’s API, generation at about one megapixel takes around 5 seconds. Larger images are much slower: the same RTX 5090 setup took about 15 seconds at 2048x2048, and the 50-step Base checkpoint is several times slower than the distilled one. These are vendor numbers on different hardware and precisions, so use them as a rough guide, not a benchmark.

Running It Locally

Running Klein yourself is realistic, but not a one-click job. You need to accept the license on Hugging Face (the repositories are gated), download the transformer, text encoder, and autoencoder, and pick a runtime: ComfyUI workflows, Hugging Face Diffusers (its ModularPipeline, currently installed from the GitHub main branch), or BFL’s reference CLI. NVIDIA also lists vLLM-Omni and SGLang Diffusion support, and Ostris AI Toolkit and NeMo AutoModel for fine-tuning.

Memory is the part to plan for. BFL says Klein runs on consumer GPUs down to an RTX 3090, and the INT8 build made with NVIDIA targets RTX cards with as little as 8 GB of VRAM. The model card’s own measurements are higher: about 31 GB peak in BF16, 26 GB in FP8, and 25 GB in NVFP4 at 1024x1024 on a data-center GPU, before the prompt upsampler, which adds about 28 GB. In practice, a home setup means a quantized build, a recent NVIDIA card (NVFP4 builds target the newest Blackwell generation), and either skipping the upsampler or running a smaller captioning model. Mac and AMD support depend on the community runtime you use; BFL’s published numbers are for NVIDIA hardware.

For fine-tuning, use Klein Base. A LoRA trained on Base also loads into the distilled Klein, so you can train slowly and generate quickly. See LoRA and quantization for the background, and when local AI makes sense for whether to bother.

API and Pricing

BFL’s API endpoint uses the same request format as the larger FLUX 3 Image and charges per image, the same for generation and editing:

ResolutionPrice per image
768x768$0.008
1K$0.009
1.5K$0.011
2K$0.014
4K$0.036

For comparison, the full FLUX 3 Image was listed at $0.048 for a 1K image at its launch. Artificial Analysis lists Klein at about $9 per 1,000 images at 1024x1024, against about $211 for OpenAI’s GPT Image 2.6 at its highest setting. Self-hosting has no per-image fee, but you pay for the hardware and your time, and larger companies need a commercial license.

Licensing

The weights ship under the FLUX Kommunity License v1.0, not an open-source license. In short:

  • Non-commercial use (research, personal projects, evaluation) is free.
  • Commercial use of outputs is free for “Qualifying Users” with less than US$5 million in gross annualized revenue, counting affiliates.
  • Everything else needs a commercial license from BFL. That includes larger companies using outputs commercially, and the model card notes the commercial-output allowance is not blanket permission for commercial fine-tuning or for redistributing the model.
  • The text encoder is Apache 2.0, but that does not change the terms for the FLUX model itself.

This is a change from the earlier FLUX.2 [klein] 4B, which was Apache 2.0, so do not assume the old terms carry over. If you are near the threshold or plan to ship a product built on the weights, read the license file and talk to BFL. The model card also lists military, surveillance, and biometric processing as out of scope. For background, see open weights.

How It Compares

On Artificial Analysis’s blind-vote leaderboards, checked November 7, 2026, Klein scored about 1043 Elo for text-to-image (24th) and about 1069 for image editing (31st). The full FLUX 3 Image sits at about 1117 for text-to-image, and OpenAI’s GPT Image 2.6 Sunburst leads at about 1259. BFL’s own charts place Klein ahead of larger open models, but those are internal preference tests without published methods. The fair summary: Klein is not the highest-quality image model, but it is among the cheapest capable ones and one of the few you can run yourself with editing and multi-reference support.

For a direct comparison with OpenAI’s hosted option, see ChatGPT Images 2.6 vs FLUX 3 Image Klein.

Not to Be Confused With

  • FLUX 3 Image, the larger BFL image model Klein keeps the capabilities of. It ranks higher on the leaderboards and costs several times more per image.
  • FLUX 3 Max, a video model that fal post-trained on BFL’s FLUX 3 foundation model. Different company, different output.
  • FLUX.2 [klein], the previous small FLUX generation, with 4B and 9B models under different licenses.
  • Hosted FLUX endpoints on other platforms, which may run different checkpoints, settings, or prices than BFL’s own API.

Strengths

  • Very low API price per image, with the same price for edits
  • Downloadable weights for private, offline, or high-volume generation
  • Editing, up to 10 references, and bounding-box layout in one model
  • A Base checkpoint built for fine-tuning and LoRA training

Limitations

  • Not top quality. Independent arenas rank it well below the leading hosted models.
  • License limits. Free commercial use stops at US$5 million in revenue, and the terms are not open source.
  • Real memory needs. The 8 GB figure applies to a specific INT8 build; the reference setup needs far more.
  • Text in images. BFL’s model card warns that text rendering can be inaccurate.
  • Prompt-sensitive. It needs detailed captions or the upsampler to follow prompts well.

For the basics, see text-to-image and diffusion models.

Continue learning

Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.

See how this tool fits into a workflow

Browse step-by-step AI workflows that use ChatGPT, Claude, Gemini, and other tools.

Frequently Asked Questions

What is FLUX 3 Image Klein?

It is the smaller, open-weights member of Black Forest Labs' FLUX 3 Image family, released October 15, 2026. A 5.6-billion-parameter image model handles text-to-image generation, image editing, up to 10 reference images, and layout control with bounding boxes. You can call it through BFL's API or download the weights from Hugging Face.

Can I use FLUX 3 Image Klein commercially?

It depends on your size. Under the FLUX Kommunity License v1.0, non-commercial use is free, and organizations with less than US$5 million in gross annualized revenue, counting affiliates, can also use the outputs commercially. Larger companies need a commercial license from Black Forest Labs or can use the paid API. The license does not automatically cover commercial fine-tuning or redistribution of the model. Unlike FLUX.2 [klein] 4B, it is not Apache 2.0.

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

Less than the big models, but more than the headline suggests. Black Forest Labs says it runs on consumer GPUs down to an RTX 3090, and an INT8 build made with NVIDIA targets RTX cards with 8 GB of VRAM. The reference code on a data-center GPU peaks around 25 to 31 GB at 1024x1024 depending on precision, and the default prompt upsampler adds about 28 GB more. Expect to choose a quantized build and possibly skip or replace the upsampler on a home GPU.

Is FLUX 3 Image Klein the same as FLUX 3 Max?

No. FLUX 3 Max is a video model that fal post-trained on Black Forest Labs' FLUX 3 foundation model; it makes video clips from text or an image. FLUX 3 Image Klein is an image model from Black Forest Labs itself, for still images and edits.

How good is FLUX 3 Image Klein compared with other image models?

Good for its size and price, but not top of the leaderboard. On Artificial Analysis's blind-vote arenas, checked November 7, 2026, it scored about 1043 Elo for text-to-image and 1069 for editing, well behind OpenAI's GPT Image 2.6 models (about 1259 and 1209) and behind the full FLUX 3 Image (about 1117 for text-to-image). It costs a fraction as much per image as those models.

Last updated: