New multimodal decisio model combi es domai specializatio , calib ated decisio s a d adaptive ve ificatio to b i g g eate p ecisio to e te p ise AI age ts
SANTA CLARA, Calif., Oct. 28, 2026 (GLOBE NEWSWIRE) — A ticul8, a p ovide of domai -specific ge e ative AI solutio s fo e te p ises, today i t oduced Bivia, a multi-modal decisio model that decides, locates, a d p oves what it sees a d ca be tailo ed to a o ga izatio ‘s i dust y k owledge, policies, a d ope ati g p ocedu es. Built o billio s of decisio s made withi A ticul8’s e te p ise AI platfo m, Bivia is ow bei g offe ed th ough a API, allowi g develope s to build AI age ts that make decisio s ac oss text, images, videos, tables a d cha ts without havi g to call a ge e ative model fo eve y step.
As e te p ises put AI age ts to wo k, they face a g owi g challe ge: the cost of usi g powe ful ge e ative models to make thousa ds of elatively simple decisio s. Acco di g to McKi sey’s State of AI i 2026 epo t, 40% of espo de ts f om la ge e te p ises said thei o ga izatio s we e scali g AI age ts, up f om 27% the p evious yea . The same esea ch fou d that app oximately o e i five espo de ts epo ted that AI- elated ope ati g costs, i cludi g toke costs, we e co st ai i g thei o ga izatio s use of AI.
U like ge e ic decisio models, Bivia ca be tu ed to a o ga izatio ‘s i dust y, policies, p ocedu es, a d histo ical decisio s, allowi g it to make judgme ts that eflect how the busi ess ope ates. It ca also evaluate text alo gside e gi ee i g d awi gs, equipme t images a d cha ts.
“We’ve bee usi g decisio models withi A ticul8 fo mo e tha th ee yea s because we saw how much u ecessa y wo k e te p ises we e aski g ge e ative models to do,” said D . A u Sub ama iya , Fou de a d CEO of A ticul8. “Whe a AI age t is wo ki g th ough a complex task, it does ‘t eed a powe ful ge e ative model to a swe eve y questio alo g the way. A d busi esses eed those decisio s to eflect thei ow k owledge a d ope ati g equi eme ts. That’s why we’ e maki g this tech ology available to eve yo e.”
I a live demo st atio , Bivia p ocessed eve y f ame of a 52-mi ute, 4K docume ta y filmed f om the I te atio al Space Statio (ISS) i a si gle equest. It t a sfo med the footage i to a sea chable i dex, p ovidi g a at-a-gla ce map of what appea ed o sc ee th oughout the film a d allowi g use s to ask questio s a d jump di ectly to the eleva t mome t. Amo g the footage we e fleeti g eve ts lasti g just seco ds, i cludi g a pass ove city lights o Ea th’s ight side a d a glimpse of the au o a. Bivia ide tified these mome ts because it a alyzed eve y f ame athe tha sampli g selected images, which could have missed them.
“Eve yo e’s models ca captio a thumb ail. Bivia watches the e ti e thi g a d tells you the th ee seco ds that matte , with the eceipts,” said D . Sub ama iya . “It’s the diffe e ce betwee a model that talks about you data a d o e that actually eads all of it.”
Bivia’s capabilities i clude:
O e model ac oss multiple modalities . Text, docume ts, images, video, audio a d aw se so sig als a e p ocessed th ough a sha ed model a chitectu e athe tha a pipeli e of sepa ate models.
St uctu ed decisio s i stead of ge e ated text . Eight ative decisio p imitives, decide, yes/ o, choose, sco e, select all, quote the sou ce, fi d the egio a d ext act the field, etu st uctu ed, machi e- eadable a swe s i a si gle pass.
Evide ce-backed a swe s . Respo ses ca ide tify suppo ti g egio s withi images, exact text passages a d cha acte offsets i docume ts, a d timestamps i video, allowi g use s to t ace decisio s to thei sou ces.
Calib ated co fide ce. Each a swe i cludes a co fide ce sco e, a d the model ca abstai whe the available evide ce is i sufficie t.
Comp ehe sive co te t a alysis . Bivia ca p ocess e ti e videos, docume ts a d audio eco di gs i pa allel, tu i g le gthy co te t i to sea chable i dexes athe tha elyi g o selected samples.
I i dust ial e vi o me ts, Bivia ca assess tech ical equi eme ts agai st suppo ti g docume ts a d images, i cludi g ide tifyi g disc epa cies betwee mai te a ce eco ds a d equipme t photog aphs. I A ticul8’s i te al testi g, Bivia achieved 100% accu acy o selected e e gy a d ha dwa e desig (Ve ilog) tasks afte bei g t ai ed o domai -specific data. These i -dist ibutio esults demo st ate its ability to be specialized fo specific i dust ial tasks, athe tha its pe fo ma ce o p eviously u see data.
I A ticul8’s i te al wo kflow completio tests i volvi g tool outi g, Bivia sco ed 0.91 agai st Jev’s 0.84. Bivia is desig ed to delive co siste t decisio s ega dless of how a swe optio s a e p ese ted, maki g it ext emely obust to cha ges a d i puts. While Jev a d othe leadi g decisio models cha ged thei a swe s i app oximately o e i eight cases, Bivia etu ed the same a swe eve y time.
Because Bivia makes decisio s without ge e ati g output toke s, develope s ca avoid usi g mo e expe sive ge e ative models fo outi e tasks, helpi g to educe the cost of u i g AI age ts at scale. The model is available th ough a API, eithe as a sta dalo e offe i g o as pa t of A ticul8’s b oade e te p ise AI platfo m.
Bivia is al eady p oduci g substa tial savi gs to A ticul8’s custome s such as Hitachi E e gy, CTC, AZi g, Megazo e a d othe s.
A ticul8 completed its Se ies B fi a ci g i Ju e 2026 at a $500 millio p e-mo ey valuatio , alo gside a st ategic pa t e ship with a global i dust ial softwa e compa y. At the time, A ticul8 epo ted mo e tha $100 millio i cumulative co t act value, with a platfo m comp isi g mo e tha two doze domai -specific AI models a d co ectio s to ove 25 e te p ise a d ope atio al systems.
