Black Forest Labs Open-source FLUX 3 Image Klein: official evaluations lead, but community hands-on results are mixed

Article

Black Forest Labs FLUX 3 Klein open-sourced: official self-tests outperform Qwen, but users find it inferior to the previous version 4B Beating AI news, Germany’s AI company Black Forest Labs open-sourced the image model FLUX 3 Image Klein. The company officially claims 5600000000 parameters, but this only counts the part responsible for image generation; in practice, it also requires 4000000000 parameters for Qwen3-VL understanding text instructions, bringing the total to approximately 9600000000 parameters. The new model natively supports 4K image generation, image editing, and up to 10 reference images. After quantization, it can run on a 8GB VRAM RTX GPU,RTX 5090 generating a one-megapixel image takes approximately 4 seconds.

Internal official evaluations show that FLUX 3 Klein’s text-to-image and image-editing capabilities both surpass Qwen-Image 2.1. Compared with its previous-generation FLUX.2 Klein 9B, the two scores increased by 91 and 70 Elo. However, initial hands-on tests in the Reddit community reported poor results. Users demonstrated failures in adjusting poses and changing viewpoints, and some even considered the results inferior to those of the previous-generation 4B model.

In another third-party blind test involving 200 real-world editing tasks,FLUX 3 Klein scored 6.93, slightly below Nano Banana’s 2.1 score of 7.20, but the cost per image was only one-fifth that of the latter. The test likewise found that it was prone to errors in pose adjustment and multi-image compositing.

Black Forest Labs particularly emphasizes that, before use, it is best to have an additional language model expand the prompt; otherwise, short instructions may produce mediocre results. To this end, the team additionally provides a 9000000000-parameter prompt expander.

The model uses the FLUX Kommunity License. Individuals and companies with combined annual revenue below 5000000 USD may commercially use generated images free of charge under the terms, but commercial fine-tuning requires separate authorization.API generating a 1K image starts at 0.009 USD.

Original link https://m.theblockbeats.info/flash/372219

Key points

Black Forest Labs open-sourced an image model FLUX 3 Image Klein。
The model officially claims 5600000000 parameters and additionally requires 4000000000 parameters to Qwen3-VL understand text instructions, for a total of approximately 9600000000 parameters.
The model natively supports 4K image generation, image editing, and up to 10 reference images.
After quantization, it can run with as little as 8GB of VRAM on a RTX GPU; reports claim that RTX 5090 generating a one-megapixel image takes approximately 4 seconds.
Official internal evaluations say that its text-to-image and image-editing capabilities surpass Qwen-Image 2.1。
In a third-party blind test,FLUX 3 Klein scored 6.93, below Nano Banana’s 2.1 score of 7.20; the cost per image was one-fifth that of the latter.
Reports mention that the model is prone to errors in pose adjustment and multi-image compositing.
The model uses the FLUX Kommunity License; individuals and companies with combined annual revenue below 5000000 USD may commercially use generated images free of charge under the terms, while commercial fine-tuning requires separate authorization.
API Generating a 1K image starts at 0.009 USD。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is an open-source model for generating and editing images. It can process high-resolution images and reference images, but its performance on complex editing tasks is disputed; when using short prompts, an additional prompt expander may be needed.

Why it matters to readers

It also offers an open-source license, a lower API starting price, and lower VRAM requirements, which may reduce the trial cost of image-generation tools; however, different tests have reached inconsistent conclusions about its performance.

Beginners should focus on points: beyond the model itself, it also requires Qwen3-VL, short prompts may produce mediocre results, and the licensing conditions for commercial fine-tuning differ from those for commercial use of images.

Risks and unknowns

The report does not provide Black Forest Labs links to the official model repository, technical report, or original test report.
The official internal evaluations, initial Reddit hands-on tests, and third-party blind-test results differ.
8GB VRAM requirements,RTX 5090 generation speed,API pricing, and commercial-use conditions still require verification against the original sources.
The specific conditions for the authorization required for commercial fine-tuning are not stated.

Black Forest Labs Open-source FLUX 3 Image Klein: official evaluations lead, but community hands-on results are mixed

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center