“We took ha dwa e most o ga izatio s would dismiss fo a model of this scale a d made it u DeepSeek V4 Flash,” said C aig K o e be ge , CEO a d co-fou de of Rival.io. “That should cha ge the co ve satio about AI i f ast uctu e a d the cost of p ivate AI. Befo e busi esses commit to mo e expe sive ha dwa e, they should ask how much mo e the ha dwa e they have could delive .”
G aphX optimizes how models a e placed ac oss GPUs a d how they use memo y a d computi g powe . This e ables e te p ises to meet wo kload equi eme ts with fewe o less expe sive GPUs a d get mo e capacity f om ha dwa e they al eady ow o ca eadily access. Depe di g o the model a d wo kload, G aphX ca educe ope -weight model hosti g costs by up to 95% compa ed with equivale t deployme ts ma aged i te ally o hosted by hype scale cloud p ovide s.
The eight L4 co figu atio combi ed 192 GB of GPU memo y a d suppo ted the model’s co text wi dow a d associated KV cache, a capability that p eviously equi ed specialized, expe sive ha dwa e setups.
Rival.io has also deployed DeepSeek V4 Flash o eight NVIDIA RTX 6000 GPUs with full e te p ise g ade pe fo ma ce, demo st ati g G aphX’s ability to suppo t e te p ise wo kloads with high co cu e cy.
“Ha dwa e specificatio s alo e do ‘t tell you what a system ca delive . How you place the model, ma age memo y, a d execute its computatio s ca fu dame tally cha ge what’s possible o that ha dwa e,” said Cha lie Wa dell, CTO a d co-fou de of Rival.io. “The implicatio s go well beyo d L4s. We’ e expa di g the ha dwa e capable of u i g la ge ope -weight models a d showi g that p ivate AI does ‘t have to depe d o a a ow set of p emium GPUs.”
G aphX optimizes wo kloads ac oss compatible AMD, I tel, a d NVIDIA ha dwa e a d suppo ts DeepSeek, GLM, a d Kimi models, givi g e te p ises flexibility to choose models a d i f ast uctu e based o thei pe fo ma ce equi eme ts, budgets, a d available capacity.
Suppo ted models ca be deployed i a e te p ise’s ow cloud o data ce te , o o dedicated Rival.io i f ast uctu e. Each deployme t ope ates i a dedicated, isolated e vi o me t with its ow secu e e dpoi t a d access c ede tials, p ovidi g co t ol ove whe e AI u s a d est icti g co ectio s to autho ized use s a d applicatio s. Models ca powe wo kflows, age ts, a d applicatio s withi Rival.io o co ect to thi d-pa ty tools, exte di g p ivate AI ac oss the systems busi esses al eady use.
G aphX is available o -dema d at Rival.io/g aph-x.
Rival.io is a e te p ise AI ope ati g system that b i gs togethe eve ythi g busi esses eed to build, deploy, gove , a d scale AI i o e i teg ated, model-ag ostic platfo m. Powe ed by Co texO e™, Rival.io e ables o ga izatio s to o chest ate AI wo kflows without ma agi g complex i f ast uctu e, while givi g develope s a d domai expe ts a way to publish, p otect, a d mo etize eusable tools, fu ctio s, age ts, a d systems. Lea mo e at ival.io.
View the o igi al p ess elease o ACCESS Newswi e
