FLUX 3

FLUX 3, launched in early access on 23 July 2026, is the point where Black Forest Labs stops being an image-model company. The German lab, founded by the researchers behind Stable Diffusion, trained a single network jointly on images, video, audio and action rather than stitching separate models behind one interface — an approach it calls Self-Flow, published in March 2026. In practice the model does four things: it generates and edits images, produces video clips of up to 20 seconds in a single pass with synchronised dialogue, sound effects and ambient audio, and predicts robot actions. That last capability is why the model belongs in a robotics catalogue at all: the same backbone that renders a scene can be asked what a manipulator should do next, and it has already been demonstrated driving robots on a production line. The caveats are large and they are the company's own doing. FLUX 3 shipped without downloadable weights and without an open licence, breaking with the FLUX.1 and FLUX.2 lines that made the lab's name — FLUX.2 [dev] is a 32-billion-parameter rectified-flow transformer with published weights for non-commercial use. Access to FLUX 3 is by request, there is no public API and no published price for any tier. Black Forest Labs has promised an open-weight FLUX 3 [dev] later in 2026. The published win rates come from the lab's own preliminary preference tests; sample sizes, rater counts and methodology have not been disclosed, and no image benchmarks were released at all. Parameter count, quantisations and hardware requirements are likewise unstated.

▸News

▸Videos

No videos yet.

▸Reviews