Black Forest Labs released FLUX 3 Image Klein on 15 October, a 5.6 billion parameter image model whose weights anyone can download and which the company says runs on an ordinary Nvidia gaming card with as little as 8 GB of video memory (VRAM). The company also claims the model ranks first among the models it compared, despite being the smallest of them. Every ranking in the announcement comes from the company’s own tests.

A person or a company can build a business on the weights at no charge only if it takes in less than $5M a year; that is the condition of the FLUX Kommunity License, which also covers research and hobby projects. Anything bigger has to pay Black Forest Labs for self-hosted licensing or use its API. The company calls the release “open weights”, meaning the trained model files are public, and that is the accurate term. “Open source” would not be, because the free tier carries a revenue ceiling.

Klein generates images from text, edits existing ones, and combines up to 10 reference images into a new picture. It outputs native 4K and accepts bounding boxes, which pin an element to a region or confine an edit to one area. The release includes a base version for fine-tuning, a distilled version (a faster variant that needs fewer steps per image), and compressed FP8, INT8 and NVFP4 builds that shrink the files. Nvidia supported those builds.

On what the post calls internal evaluations, Klein takes first place for editing images, for generating from reference pictures, and for plain text prompts, ahead of bigger models with public weights. Two closed models in the same lineup, FLUX.2 [pro] and Ideogram V4.5, score lower. Its rivals are Qwen-Image 2.1 at 7 billion parameters and Krea 2 at 12.9 billion. Against the company’s previous FLUX.2 [klein] 9B, the company reports a jump of 91 Elo points on text prompts and 70 on editing, while using 38 percent fewer parameters. Elo is a rating borrowed from chess, where higher means the model wins more head-to-head matchups; the post does not name an outside evaluator or describe who judged. Krea 2 cannot edit images, so it sits out the editing comparison. The parameter counts leave out text encoders, the separate components that read the prompt.

Full BF16 precision needs 11.2 GB for the generator weights, and FP8 halves that to 5.6 GB. Pair the distilled model with the INT8 build and a 1 megapixel image takes roughly 4 seconds on an RTX 5090, or just under 9 on an RTX 4070, both in ComfyUI.

Those timings carry a footnote from the publisher: they were measured by Nvidia. The chip vendor also helped build the compressed versions, and Black Forest Labs trained the model on Nvidia GB200 NVL72 systems. The chip vendor has a stake in showing its cards handle the workload, so read the figures as vendor demonstrations, not independent tests.

The speed comparison against rivals tilts toward the publisher by its own description. Through the paid API, Black Forest Labs says Klein answers faster than Krea 2, Ideogram V4.5, FLUX.2 [pro] and Qwen-Image 2.1. It timed its own model on its own servers, while it timed the others from the client side, which includes polling and network delay. Krea 2 was measured on its Medium tier; the post adds that the Large tier needed 31.1 seconds on a text prompt.

On cost, the API charges from $0.009 per 1K image, and the distilled model is sold there as flux-3-image-klein. Weights for the base and distilled versions are on Hugging Face. From day one it works with vLLM-Omni, SGLang Diffusion, TensorRT-LLM, Diffusers from Hugging Face and ComfyUI, and serving a custom LoRA (a small add-on that teaches a model a style or subject) adds about 5 percent to generation time. On safety, the company says it red-teamed the checkpoints internally and that its external partner Cinder found substantially fewer vulnerabilities than in other popular open-weights models. That is the company’s account of Cinder’s work, not a published report.

FLUX 3 Action came first; Klein is the family’s second release with public weights. For a studio or startup under $5M in revenue, the cost of trying it is a download and an afternoon; a larger company should price the self-hosted licence before building a pipeline on it, and should run its own side-by-side images before believing any ranking.

Based on the Black Forest Labs release post “FLUX 3 Image Klein” of 15 October 2026.