One model family shipped under three different licences

The process wall

On 1 August 2024, a new company released a model family built by researchers behind the architecture much of open-source image generation already ran on. The launch announcement describes the founding team's prior work as including VQGAN, Latent Diffusion, Stable Diffusion models for image and video generation, and Adversarial Diffusion Distillation, the same lineage this batch's other diffusion notes trace. The process wall FLUX.1 addressed was licence fragmentation: rather than one release under one licence, Black Forest Labs shipped three tiers with three different terms attached to the same underlying model family.

What the documents show

The announcement names three variants: FLUX.1 [pro], available only through an API; FLUX.1 [dev], described as an open-weight, guidance-distilled model for non-commercial applications, with commercial use requiring separate contact with the company; and FLUX.1 [schnell], openly available under an Apache 2.0 licence. The FLUX.1-dev model card confirms it is released under the FLUX.1 [dev] Non-Commercial Licence specifically, describes the model as a 12 billion parameter rectified flow transformer, and lists prohibited uses including exploiting minors and creating non-consensual intimate imagery, without stating what the model was trained on. The schnell model card confirms the Apache 2.0 licence and states the model was trained using latent adversarial diffusion distillation to generate high-quality images in only one to four steps, a distillation method, not simply a smaller model.

Authorship and limitation

The three-tier structure means the same architecture carries three different sets of rights depending on which weights a maker downloads, which is the mechanism a licence-only summary such as FLUX is open source obscures: schnell's Apache 2.0 terms are genuinely permissive, dev's non-commercial terms are not, and pro is not open-weight at all. Neither model card discloses training data, so a maker cannot check the training-data question here the way the Stable Diffusion or Adobe Firefly records allow; the absence of that disclosure is itself worth noting rather than filled in with assumption. The distillation method behind schnell's speed is a training-time technique, not evidence about what the model can or cannot generate relative to the slower variants.

What survives

What a maker must check before use is which specific variant and licence a downloaded checkpoint carries, since the model family name alone does not indicate the terms. Editorially, describing FLUX as open source without naming the variant risks misrepresenting the dev and pro tiers' actual restrictions.

• Confirm which FLUX.1 variant, pro, dev or schnell, a specific tool or checkpoint uses before citing its licence.
• Read the dev model card's non-commercial terms in full before any client-facing use.
• Check whether a training-data disclosure has since been added to either model card, since neither states one as retrieved on 16 September 2026.

Black Forest Labs' announcement is best read as a licensing map rather than a single release: the technical lineage is continuous with earlier open diffusion work, but the rights attached to each variant are not, and that distinction sits entirely with the documents, not with the model's outputs.