Black Forest Labs releases FLUX 3 Image Klein as open weights

On 15 October 2026, Black Forest Labs published FLUX 3 Image Klein. The post calls it a 5.6B-parameter open-weights model for text-to-image generation, instruction editing, and multi-reference generation with up to 10 reference images. The bfl_ai post at 15:14 UTC the same day repeats the 5.6B figure, native 4K, bounding boxes, and what it calls pixel-perfect editing. It is the second open-weights release in the FLUX 3 family, after FLUX 3 Action, covered here as a robot-control model in the 24 September account. Krea’s hosted image API is a different product, already reported in the Krea 3 account.

Two checkpoints, and a gated Hub repo

The API docs describe two variants. Distilled is step-distilled, 16 inference steps, guidance 1, for fast generation. Base is undistilled, 50 steps, guidance 3.5, for fine-tuning, LoRA training, and research. Both are labeled 5.6B, and the docs list a maximum resolution of up to 4K. Native 4K is the vendor’s claim on the blog and on the bfl_ai post.

The distilled checkpoint is black-forest-labs/flux-3-image-klein. The base checkpoint is black-forest-labs/flux-3-image-klein-base. On 16 October the Hub API records for the distilled repo and the base repo listed both as public, with gated set to auto, license set to other, and license_name set to flux-kommunity-license. The model pages say you have to accept conditions, including sharing contact information, before the files open. The distilled repo’s createdAt is 13 October 2026 and its lastModified is 16 October 2026. The blog date is 15 October.

The distilled card says the 5.6B label is the image transformer. Qwen3-VL-4B-Instruct is a separate text encoder, and the card says that encoder is under Apache-2.0. The image weights are under the Kommunity license described below.

The card says boxes in the prompt place elements on a 0 to 1000 grid. They guide placement and scale. They do not constrain pixels to a hard mask.

The blog says the release includes FP8, INT8, and NVFP4 builds developed with NVIDIA. The distilled Hub file list includes variants/fp8r and variants/nvfp4. It does not include an INT8 directory.

The docs say not to run Klein without a prompt upsampler. The card names black-forest-labs/flux-3-image-klein-pup and says a short prompt sent as typed gives generic output. Diffusers does not call the upsampler on its own. The card says the upsampler adds about 28 GB of GPU memory.

The Kommunity license, as the card states it

The card says the weights use the FLUX Kommunity License v.1.0: non-commercial use, plus commercial output generation by Qualifying Users with less than US$5M in gross annualized revenue, including affiliates, subject to the license terms. It also says that commercial exception is not a blanket permission for commercial fine-tuning or distribution. The raw LICENSE.md URL refused access without a logged-in acceptance of the gate, so this paragraph follows the card and the blog, not a direct read of the license file.

The blog’s shorter line is free commercial use for individuals and companies with annual revenue under $5M, alongside research and non-commercial use. Larger organizations, it says, can deploy commercially through self-hosted licensing or the API. The licensing page is a paid self-hosted menu that includes FLUX 3 Image Klein. Those paid tiers are a different grant from the Kommunity terms on the card. The image weights are not an Apache or MIT release. The Apache-2.0 line on the card applies to the Qwen text encoder.

API prices on the docs

The docs price text-to-image and editing the same, per output image: $0.008 at 768 by 768, $0.009 at 1K, $0.011 at 1.5K, $0.014 at 2K, and $0.036 at 4K. The blog says pricing starts at $0.009 per 1K image, which matches the 1K row and leaves the $0.008 row unstated. The docs say the API serves the distilled model as flux-3-image-klein. Base is listed as Hugging Face only.

Ranks and timings are the vendor’s

In internal evaluations, the blog says Klein ranks first in text-to-image, multi-reference generation, and image editing among the open-weights models it compared, and above the closed models in that comparison, Ideogram V4.5 and FLUX.2 [pro]. It says Klein is the smallest open model in the set, at 5.6B against 7B for Qwen-Image 2.1 and 12.9B for Krea 2, and that it gains 91 Elo in text-to-image and 70 in image editing versus FLUX.2 [klein] 9B, with 38 percent fewer parameters. Those Elo figures and the ranking are Black Forest Labs’ internal evaluation.

The blog prints generation times and says NVIDIA measured them. With the distilled model and the INT8 build, it says about 4 seconds for a 1 megapixel image on an RTX 5090 and under 9 seconds on an RTX 4070, through ComfyUI. On an NVIDIA GB300, it says the distilled model generates a 1 megapixel image in under 0.6 seconds in NVFP4.

The blog says day-one support includes ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni, plus LoRA training on the base checkpoint in Diffusers and Ostris’s AI Toolkit.