FLUX 3 Image Klein is out as open weights, and OpenMOSS released MOSS-Image a day later

Black Forest Labs says Klein runs in 8GB, but its own model card measures 25GB or more and wants a 9B prompt writer in front.

By Think Facility · October 19, 2026 · Black Forest Labs Open weights

Black Forest Labs released FLUX 3 Image Klein on October 15, 2026. It's a 5.6B-parameter image model that makes pictures from text, edits them and combines up to 10 reference images, and the weights are on Hugging Face in two versions: a base checkpoint for fine-tuning and a distilled one for generation. Distilled here means it was trained to finish an image in 16 sampling steps where the base model takes 50.

A day later, on October 16, OpenMOSS published the weights for MOSS-Image, a 5B image model of its own, in a Base and an SFT version. Its GitHub page puts the main pitch on text rendering and layout control.

Klein: 5.6B parameters, open weights
Klein license: FLUX Kommunity License
MOSS-Image: 5B backbone
MOSS-Image license: Apache License, Version 2.0

What Black Forest Labs says Klein can do

Three men with blue conference lanyards sit at a black table in a pale wood-paneled room, facing a woman in a pinstripe jacket. — European Commissioner Henna Virkkunen (right) meeting Black Forest Labs at the European Digital Sovereignty Summit in Germany, November 18, 2025. Photo: Europäische Kommission - Audiovisueller Dienst, CE - Service audiovisuel, Stefanie Loos, EC - Audiovisual Service, FACTSTORY, CC BY 4.0, via Wikimedia Commons

BFL says Klein "shares core capabilities with the full FLUX 3 Image model", including native 4K output and bounding boxes for placing things or fencing an edit into one area. The bigger FLUX 3 Image model isn't among the downloads. BFL lists two open releases from the FLUX 3 family so far: FLUX 3 Action, a 7B model for robot control, and now Klein.

Every quality number in the launch post is from BFL's own internal evaluations. By those, Klein ranks first among open models in all three of its jobs and rates above two closed ones, Ideogram V4.5 and FLUX.2 [pro]. Against BFL's previous small model, FLUX.2 [klein] 9B, it gains 91 Elo in text-to-image and 70 in editing, with 38% fewer parameters. It's also the smallest model in the comparison: Qwen-Image 2.1 is 7B and Krea 2 is 12.9B, counting the image generator only.

8GB on one page, 25GB and up on another

The hardware figures depend on which page you read. The announcement says the INT8 build runs on GeForce RTX cards with as little as 8GB of VRAM through ComfyUI or Diffusers, and that a 1MP image takes about 4 seconds on an RTX 5090 and under 9 on an RTX 4070.

The model card measures something else. With BFL's reference code on a B200, a data-center GPU, at 1024x1024 and with the prompt upsampler switched off, peak memory is 31 GB in BF16 and 25 GB in NVFP4. I read those as two different setups, though neither page explains the gap.

The upsampler matters more than it sounds. Klein expects a full scene caption (lighting, camera, where each thing sits), and the card says not to run it without something to write one:

a short prompt sent as typed gives generic output.

From black-forest-labs/flux-3-image-klein · Hugging Face

BFL's own upsampler is a separate 9B model, fine-tuned from Qwen3.5-9B and released under Apache 2.0, and its weights are 18.8 GB in bfloat16. The card says it adds about 28 GB on an 80 GB GPU. Any other LLM can stand in if you give it BFL's system prompts, which is the cheaper route on a home machine.

MOSS-Image

OpenMOSS describes MOSS-Image as a 5B base DiT model. It takes prompts of up to 4K tokens, and you can hand it a JSON layout instead of a sentence, giving each piece of text and each object its own box in pixel coordinates. The examples on its card render handwritten Chinese on a sticky note, and the bundled VAE uses the FLUX.2 small decoder.

There are no benchmark scores on its Hugging Face card or GitHub page. The one comparison is a single example where the SFT version lays out a whiteboard more neatly than Base. The technical report went up on arXiv on October 19.

Who can use them for paid work

For anyone thinking of putting these into a product, the licenses differ more than the models do. Klein's FLUX Kommunity License allows free commercial use for individuals and companies with annual revenue under $5M, plus research and non-commercial use. Bigger companies need BFL's self-hosted licensing or its API. The card also lists uses the license rules out, among them surveillance, biometric processing and using Klein's outputs to improve a competing image model.

MOSS-Image's weights and inference code are under Apache 2.0, which grants a "perpetual, worldwide, non-exclusive, no-charge, royalty-free" copyright license with no revenue line at all. So a freelance illustrator or a small studio can use either one commercially today. A company over $5M can self-host only MOSS-Image without a separate deal.

BFL's post ends by saying more FLUX 3 open releases will follow, and gives no date.