FREIBURG, Ge ma y, July 23, 2026 (GLOBE NEWSWIRE) — Black Fo est Labs, the global f o tie AI esea ch lab buildi g the fou datio laye fo visual i tellige ce, today i t oduces FLUX 3, its ew multimodal f o tie model. FLUX 3 joi tly lea s f om images, video, a d audio withi a u ified a chitectu e, a d ca also be exte ded to p edict actio s. The model is Black Fo est Labs’ ewest additio to its FLUX family of visual AI models, which a e k ow fo pai i g state-of-the-a t capabilities with ope access, a d fo powe i g ge e ative featu es i side leadi g c eative, develope , a d co sume platfo ms like Adobe Photoshop, Picsa t, Nous Resea ch’s He mes Age t, a d mo e.
The u veili g of FLUX 3 comes at a i flectio poi t fo the i dust y, as the focus moves beyo d la guage models towa d i tellige t systems that ca u de sta d both the visual a d physical wo ld. While this shift has p oduced a cluste of adjace t fields—image a d video ge e atio , wo ld models, simulatio , obotics a d physical AI—Black Fo est Labs sees these catego ies as i te co ected exp essio s of the same u de lyi g capability: visual i tellige ce, o models that ca pe ceive, p edict, a d act ac oss physical a d digital e vi o me ts.
“We place visio at the ce te of ou app oach because it is the most sig al- ich medium of the physical wo ld. Images co vey st uctu e, images a d video teach spatial elatio ships, video teaches dy amics, a d actio s eveal causal elatio ships. But visio alo e is ot the complete pictu e,” said Robi Rombach, Co-Fou de a d CEO of Black Fo est Labs. “T ue i tellige ce mea s pe ceivi g the wo ld: p edicti g how it will cha ge, taki g actio , a d lea i g f om the esults. Joi t t ai i g withi o e u ified a chitectu e is what will get us the e, because each t ai i g modality st e gthe s the othe s. Audio co veys timi g, p osody, a d physical eve ts that elude visio . La guage co veys goals, abst actio s, a d i st uctio s that pixels ca ot easily exp ess.”
“You ca ’t cheat eality. A model that o ly lea s images ca o ly ge e ate images. But the wo ld is ot made of still f ames. It moves, sou ds, cha ges, a d espo ds,” said Rombach. “That’s why FLUX 3 is t ai ed ac oss those sig als togethe , so the model ca build a deepe u de sta di g of how the wo ld wo ks. That is what visual i tellige ce equi es, whethe the applicatio is video ge e atio , simulatio , o obotics.”
FLUX 3 is built to suppo t applicatio s i cludi g i c eative tooli g, media, desig , e-comme ce, a d physical AI—i cludi g video ge e atio with sy ch o ized audio, p ecisely edited images, the mai te a ce of p oduct a d mate ial co siste cy ac oss motio , a d actio p edictio fo obotics.
The model will be available th ough FLUX 3 Video, FLUX 3 Image, FLUX 3 Actio , a d FLUX 3 Dev. FLUX 3 is al eady bei g tested by Ca va, Bu da, Mag ific (fo me ly F eepik), K ea, a d Picsa t. While still i developme t, FLUX 3 Video al eady leads i ea ly evaluatio s agai st f o tie video models, a d is pa ticula ly st o g i captu i g huma facial exp essio s, associati g sou ds with physical eve ts, a d multili gual capabilities.
FLUX-mimic is desig ed fo ge e al-pu pose obotic ma ipulatio : helpi g obots u de sta d a visual sce e, p edict the co seque ces of a actio , a d adapt to ew tasks with fa less task-specific data. Depe di g o task difficulty, the model ca be fi e-tu ed fo a specific ma ipulatio task with as little as 30 mi utes of obot data, whe e p io app oaches have equi ed 30 o mo e hou s.
“The ha dest pa t of obotics is data,” said Elvis Nava, CTO of mimic obotics. “Eve y ew task o mally mea s hou s of a obot epeati g itself. Because FLUX-mimic is built o top of f o tie video models that al eady u de sta d how the physical wo ld behaves, it picks up a ew task i mi utes, ot days. This way, we ca leapf og the cu e t state of the a t i obot lea i g.”
“I pa t e ship with mimic, Audi has bee testi g a d deployi g FLUX-mimic. We have see these obots solve complex soft body ma ipulatio wo k that would have bee simply impossible with co ve tio al obotics. This ca have a majo impact i assisti g ou employees, i c easi g efficie cy, a d expa di g flexible automatio ac oss p oductio a d logistics ope atio s,” said Ch istoph Sch eide , Audi P oductio Lab. “Fo us, pa t e i g with pio ee i g compa ies such as mimic a d Black Fo est Labs is esse tial i pushi g the f o tie of Physical AI a d validati g these i ovatio s i eal-wo ld p oductio e vi o me ts.”
Black Fo est Labs will also elease faste a d ope -weight ve sio s of FLUX 3 late this yea , co siste t with the compa y’s lo g-held belief i buildi g a d sha i g its tech ology ope ly so esea che s a d develope s a ou d the wo ld ca build, test, a d deploy ew applicatio s o top of it. Ope weights make secu e, low-late cy local deployme t possible fo applicatio s like obotic co t ol systems, a d let teams adapt FLUX 3 to thei ow data, p oducts, a d wo kflows—tu i g a ge e al model i to a secu e fou datio fo a y c eative, simulatio , o physical AI stack.
“We a e o ly begi i g to sc atch the su face of ve satile, capable, u ified visual models,” said Rombach. “F om i te active image a d video editi g to simulatio , physical AI, a d compute use, the f o tie is wide ope .”
Black Fo est Labs has quickly become a sta da d visual AI e gi e ac oss the builde a d c eative ecosystems si ce its 2024 lau ch f om stealth. Its fou de s have bee pio ee i g mode visual AI fo fa lo ge , f om VQGAN, late t diffusio , a d Stable Diffusio to Black Fo est Labs’ family of FLUX models, which powe AI capabilities fo leadi g e te p ise platfo ms a d c eative i dust y p ofessio als like lege da y film di ecto Ma ti Sco sese. To date, models f om the Black Fo est Labs fou de s have bee dow loaded ove half a billio times.
