Germany's FLUX 3 Klein image generator runs on a PC, but it isn't free for everyone
On October 15, Black Forest Labs released the weights of FLUX 3 Image Klein, a 5.6-billion-parameter model. Here is what it promises on a consumer graphics card, and who has to pay for a license.
Translated from French by AI. The original French article is the reference version.

Eight gigabytes of video memory: that is the minimum announced by Germany's Black Forest Labs to run its new image generator on an NVIDIA GeForce graphics card. On October 15, the Freiburg-based company released FLUX 3 Image Klein, an “open-weight” model: you can download its parameters and run it at home, without going through an online service.
The model has 5.6 billion parameters. It can create an image from text, edit an existing image and combine up to ten reference images. Black Forest Labs also highlights native 4K generation and the ability to outline, with a box, the area of an image to edit. It is the second open model in the FLUX 3 family, after FLUX 3 Action, unveiled in September and aimed at robotics.
What hardware do you need?
Several versions are available: a base model, meant to be retrained, a faster “distilled” version, and compressed versions (FP8, INT8, NVFP4) developed with NVIDIA. The weights take up 11.2 GB in BF16 and 5.6 GB in FP8. The INT8 version is the one meant to fit in 8 GB of memory.
On speed, the company claims about 4 seconds for a one-megapixel image on an RTX 5090 and under 9 seconds on an RTX 4070, using the distilled version. These figures were measured by NVIDIA and published by Black Forest Labs. VentureBeat points out that the 8 GB configurations and 4K do not appear in the published charts, and that a 2048 × 2048 image takes noticeably longer: 50 seconds on an RTX 4070, according to the company's tables.
“The best open model”: according to whom?
Black Forest Labs says Klein ranks first among open models for image generation and editing, and ahead of closed models such as Ideogram V4.5. These results come from its own blind evaluations. VentureBeat notes that the charts give little detail on the prompts used, how results were judged and the sample sizes. No independent evaluation is available at this stage. The same caution applies to security: the company says its partner Cinder found “significantly fewer vulnerabilities” than in other open models, without publishing any figures.
Open, but not free for everyone
This is the point to check before any professional use. The model is released under the FLUX Kommunity License, not an open-source license. Commercial use is free for individuals and for companies with annual revenue under $5 million. Above that, you need a paid hosting license, whose price is not published, or you must use the company's API (application programming interface). The API is priced from $0.009 per image at 1K resolution, according to VentureBeat. Non-commercial use and research remain free.
For an SME, an agency or a communications department, Klein therefore offers an interesting option: generating visuals locally, without sending your images to a US service. A mid-sized company will have to negotiate. As with NVIDIA's language model that fits on a single graphics card, the real test will come from independent evaluations. They will show whether the announced performance holds up on an ordinary machine.
Method: AI-assisted monitoring, verified sources, article reviewed and approved by Frédéric LENQUETTE.
Share this article