Black Forest Labs came out of the original Stable Diffusion team and it shows in the documentation, which is written by people who understand the models rather than by a marketing function describing them. The coverage spans the FLUX family across versions, the Kontext editing models, FLUX Tools, the API, self hosting and an MCP integration, and it is organised so that you can find the thing you need without reading the rest. The company raised a three hundred million dollar Series B at a three and a quarter billion valuation in December 2025, and unlike several of its peers it has kept shipping open weights alongside the hosted API. The editing material is the strongest part.
Kontext made instruction based image editing work properly, meaning you can describe a change to an existing image and get that change rather than a new image that vaguely resembles the old one, and the guides teach how to phrase those instructions well. That is real craft knowledge and it is specific, with worked examples of instructions that work and instructions that do not. The prompting material is similarly grounded, teaching structure and composition rather than reciting incantations, which is the failure mode of almost every prompting guide written for image models. Documenting self hosting seriously is the other thing I want to credit, because open weights are only meaningfully open if somebody tells you how to run them, and plenty of labs release weights and then quietly make the hosted path the only documented one.
My reservations are commercial and ethical rather than technical. Licensing across the FLUX variants is genuinely confusing. Different models in the family carry different terms, some permit commercial use freely and some do not, and working out which situation you are in requires more care than the documentation makes easy. For anybody planning to ship a product on these models that is the most important question they have, and it deserves a clear comparison table rather than terms scattered across model pages.
Pricing has a similar problem, with video priced transparently at around seventeen cents a second and per image rates harder to pin down. On training data there is nothing. That is now standard across the industry and it does not make it acceptable, particularly for a tool sold to people who will use the output commercially and who carry the downstream risk of whatever the model absorbed. The absence of any guidance on disclosure or provenance marking is the other omission, and for a lab this technically serious it feels like a deliberate silence rather than an oversight.
Three point five. The most technically credible documentation in image generation, with genuinely useful editing craft, marked down for confusing licences and for saying nothing about where any of this came from.