Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights image model

What Black Forest Labs released on 15 October

Black Forest Labs, the German startup behind the FLUX image models, published the release of FLUX 3 Image Klein on 15 October 2026. It is a 5.6-billion-parameter open-weights model for image generation and editing, and the company calls it the second open-weights release in the FLUX 3 family, after FLUX 3 Action, its robot-control model.

The weights are on Hugging Face, the reference inference code is on GitHub, and the distilled model is also served through the company's own API. 'Klein' is Black Forest Labs' name for the smaller, more efficient members of each FLUX family, meant to run on hardware a lot of people actually own rather than only in a data centre.

The release lands the same day as a cluster of image-tooling news, including Krea's Krea 3 editor beta, so it is worth being precise about what Klein is: an open-weights model you download and run, not a hosted service. That is the part that makes it interesting for people who want to keep image work on their own machines.

What FLUX 3 Image Klein does

Klein does three jobs in one set of weights. It generates an image from a text prompt; it edits an existing image from an instruction; and it composes a new image from up to 10 reference images, so a character, a product or a style can be carried across pictures. Black Forest Labs says it shares core capabilities with the full FLUX 3 Image model, including native 4K generation and what it calls pixel-perfect editing, where unedited parts of an image are preserved exactly.

It also takes bounding boxes. You can give a region of the canvas and a short description, and the model places that element inside the box, on a 0-to-1000 grid that holds whatever the aspect ratio is. The company is careful to say the boxes guide placement and scale rather than acting as a hard mask: an element can spill past its box, small requested elements may not appear, and an edit can change content outside the region you asked about. That honesty about limits is unusual in a launch post.

One practical catch from the model card: the model is trained on full scene captions written by a prompt upsampler, and Black Forest Labs says not to use it without one. A short prompt typed as-is gives noticeably weaker, more generic results. The reference code loads the company's upsampler by default. This is the same lesson that keeps surfacing with modern image models, including in our own test of Google's Nano Banana 2.1: prompting style changes the output as much as the model does.

What it takes to run it

Klein ships in several forms. There is a distilled checkpoint built for fast generation and editing, which runs in 16 steps without classifier-free guidance, and a base checkpoint built for fine-tuning, LoRA training and research, which runs in 50 steps. On top of those, Black Forest Labs released FP8, INT8 and NVFP4 builds developed with NVIDIA, so the model can be quantised down to smaller hardware.

The headline for home users is the INT8 build: Black Forest Labs says it runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM, through ComfyUI or Hugging Face Diffusers. The generator weights are 11.2GB in BF16 and 5.6GB in FP8. On a vendor-measured benchmark, the distilled model makes a one-megapixel image in about four seconds on an RTX 5090 and under nine seconds on an RTX 4070 through ComfyUI, and on an NVIDIA GB300 the NVFP4 build gets under 0.6 seconds. Those timings are the GPU's own generation time and were measured by NVIDIA, not independently.

There is day-one support in ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni, and LoRA training through Ostris's AI Toolkit and NVIDIA's NeMo AutoModel. The small size is the point: fewer parameters make the model cheaper to adapt and realistic to run locally, which matters for teams that need to keep data on their own machines. It fits the same shift toward on-device model running we have tracked with tools like Hugging Face's llama.cpp Windows tray app.

What the FLUX Kommunity License allows

The weights come under the FLUX Kommunity License v1.0. In the company's own summary, it allows non-commercial use, plus commercial generation of outputs by what it calls Qualifying Users: individuals and companies, including affiliates, with less than US$5M in gross annualised revenue. Larger organisations that want to deploy commercially need the company's self-hosted licensing or its API. The Qwen3-VL-4B-Instruct text encoder that Klein uses is separately under Apache-2.0.

The license is not blanket permission. Black Forest Labs says its commercial-use exception does not by itself cover commercial fine-tuning or distribution, and the model card points to a list of out-of-scope uses in the license: military purposes, surveillance, biometric processing, generating unlawful content including child sexual abuse material or non-consensual intimate imagery, violating rights of publicity or digital replica rights, and using the model or its outputs to improve another image model that competes with FLUX. Export controls and sanctions apply too.

That competitor clause is worth noting for anyone planning to train on Klein's outputs: the license specifically forbids using the model, its derivatives or its outputs to improve another model that does a similar job. Black Forest Labs also says use of the FLUX and Black Forest Labs trademarks needs prior written permission beyond ordinary attribution. In short, it is open weights with real strings attached, not a permissive licence, and anyone building a product on it should read the actual terms.

What Black Forest Labs claims, and what it hasn't shown

Black Forest Labs' quality claims are from its own blind evaluations. It says Klein ranks first among open-weights models in text-to-image, multi-reference generation and image editing, ahead of larger open models, and that it rates above the closed Ideogram V4.5 and the company's own FLUX.2 pro in the same comparison. It puts Klein at 5.6B against 7B for Qwen-Image 2.1 and 12.9B for Krea 2, and reports gains over FLUX.2 klein 9B of 91 Elo in text-to-image and 70 in editing with 38% fewer parameters.

Those are Elo scores from internal preference tests, not an outside benchmark, and the parameter counts cover the image generator only, not the separate text encoder. The company also measured its latency against rivals partly on its own servers and partly from a client, which it notes is not a like-for-like setup. None of that makes the numbers wrong, but they are the vendor's, and AI Lookout has not run the model.

On safety, Black Forest Labs says it filtered training data, fine-tuned to suppress harmful behaviour, red teamed the checkpoints internally, and had an external partner, Cinder, evaluate them, reporting substantially fewer vulnerabilities than other popular open-weights models. As with the quality figures, that is the company describing its own process; the useful independent signal will come from whatever the open-source community finds once the weights are in wide use.

Where Klein sits among open image models

Klein lands in a crowded month for open image generation. It is explicitly pitched against Qwen-Image 2.1 and Krea 2 on size and quality, and it follows a run of small, local-first model releases we have been tracking, from Sperid Labs' 3B pixel-space Iris model to the broader wave of on-device models. The common thread is shrinking a capable model until it fits a single consumer GPU.

The distinctive thing about Klein is breadth in one checkpoint: generation, instruction editing and multi-reference composition without swapping models, plus bounding boxes for layout, at a size that runs on 8GB. Most open models do one or two of those well. Whether Klein actually beats the field on quality is the open question, and it is one the community can now answer directly, because the weights are downloadable rather than locked behind an API.

For the home or small-studio user, the calculus is simple. If you have an RTX card and want generation and editing on your own machine without per-image fees, Klein is now one of the strongest things you can run locally, within a licence that is free for small commercial use. If you are a larger company, the licence pushes you toward the API or a paid self-hosted deal, which is where Black Forest Labs makes its money.

What to keep in mind

The quality and speed numbers are the vendor's, from internal Elo tests and NVIDIA-measured timings, so treat the 'best open model' framing as a claim to be checked rather than a settled result. The 8GB figure is for the INT8 build in ComfyUI; the reference code has no INT8 path and uses more memory, and the prompt upsampler the model needs adds a substantial amount of GPU memory on top.

The licence is the other thing to read carefully. 'Open weights' here means downloadable and usable within limits, not permissive: the US$5M revenue cap, the ban on training a competing model from Klein's outputs, and the trademark restrictions are all real. For a product plan, those terms matter as much as the benchmark table.