Black Forest Labs releases image-model weights with limited free commercial rights

Qualifying smaller businesses can now generate commercial images on their own hardware without paying Black Forest Labs a model-license fee. The company released FLUX 3 Image Klein weights on October 15, giving developers a local alternative to its hosted service. But the revenue-based exception covers commercial outputs—not unrestricted rights to fine-tune or distribute the model.

The $5 million line includes affiliates

The FLUX Kommunity License permits noncommercial use and commercial output generation by qualifying users with less than $5 million in gross annualized revenue. Eligibility includes affiliates—not just the budget of a particular design team or pilot project. That makes the corporate relationship relevant even when the proposed deployment is small.

Black Forest Labs’ model card explicitly warns that this exception is not blanket permission for commercial fine-tuning or distribution. Access also requires users to accept conditions and share contact information. The repository is publicly visible, but downloading the files is a separate step governed by those terms.

Larger organizations need separate terms for commercial self-hosting or can buy API access. VentureBeat reports BFL’s published prices as $0.009 for one 1K-resolution image, $0.014 at 2K and $0.036 at 4K. Generation and editing cost the same at each resolution. BFL has not published a standard commercial self-hosting fee.

Two versions for different jobs

Klein combines text-to-image generation, edits to existing pictures and compositions using up to ten reference images. Its advertised 5.6 billion parameters count the image-generating component; the text encoder is separate. BFL also says Klein supports native 4K output, although its supplied local speed tests do not measure that resolution.

• Base: the downloadable starting point for fine-tuning, lightweight adaptation known as LoRA training, and research. Its standard configuration uses 50 generation steps. Technical support for adaptation does not override the license’s commercial restrictions.
• Distilled: a version designed for generation, editing and self-hosting, using 16 steps instead of 50. This is also the checkpoint behind BFL’s API, rather than the base version intended for training workflows.

Developers can describe where objects should appear using bounding boxes: coordinates that guide placement and scale. Those boxes are not hard pixel masks. BFL warns that objects can extend beyond them, small requested elements may disappear, and edits can change content outside the requested region. Reference images anchor content without locking the composition.

Prompt style also affects the result. BFL recommends full descriptions of the intended scene, including framing, subjects, background and lighting. Its reference software can run a separate prompt-expansion model that rewrites short requests into detailed captions. Without that assistance, BFL warns that short prompts can produce generic images and brief editing instructions can reframe the shot.

An 8 GB minimum is not a speed promise

BFL specifies an 8 GB video-memory minimum for local use on GeForce RTX hardware. It has not attached that minimum to a particular variant, output resolution or generation time. The base model card, meanwhile, reports 26–31 GB of peak GPU memory for its reference implementation on a B200 at 1,024 × 1,024, with prompt expansion disabled.

The company’s speed measurements show how sharply the workload changes with resolution. On an RTX 4070, its 16-step INT8 build generated a 1,024 × 1,024 image in 8.9 seconds, rising to 50.4 seconds at 2,048 × 2,048. Editing took 18.9 seconds and 150 seconds, respectively. These are company-supplied results, not independently reproduced measurements.

BFL’s internal preference tests also place Klein ahead of the downloadable-weight alternatives shown in its charts. But the company has not supplied enough detail about prompts, judges, sample sizes or uncertainty to establish the size of that advantage independently. The local speed tests supply no matching competitor runs or quality comparison between the 16-step and 50-step versions.