Black Forest Labs unveiled FLUX 3, a multimodal flow-matching model released July 23, 2026, that generates video up to 20 seconds with synchronized audio alongside image synthesis and action prediction. The model employs ‘Self-Flow,’ a self-supervised training framework that jointly optimizes generation and representation quality, reportedly improving manipulation task success rates from 42% to 71%. In human preference evaluations, FLUX 3 achieved a 93% preference rate over Luma Ray 3.2, though its open-weight variant, FLUX 3 Dev, remains unreleased with current access limited to early adopters.
