The cost p oblem is st uctu al. The flat- ate subsc iptio pla s that make popula codi g age ts affo dable fo i dividual develope s come u de co sume te ms, whe e a p ovide may use p ompts a d code to imp ove its models. Regulated e te p ises ca ot se d p op ieta y code, custome eco ds o vehicle data th ough those te ms, so they a e left buyi g eve y toke at comme cial API ates. As age t swa ms scale, so does the bill.
MetaLLM attacks that bill f om two sides. Fi st, it acts as a dy amic p oxy laye fo the whole age t swa m, weighi g the complexity of each task a d b oke i g it to the model best suited to it. Routi e wo k such as pa si g, t acki g updates a d status checks goes to lightweight models hosted o the compa y’s ow ha dwa e, a d o ly the ha dest tasks each p emium f o tie models. Seco d, it gets cheape the lo ge it u s. Pai ed with RuVecto self-lea i g memo y, MetaLLM sto es eve y task its age ts complete successfully a d, just as impo ta t, eve y patte that failed. Each outi g decisio d aws o that histo y, so MetaLLM ca ha d mo e of the wo k to less expe sive models with co fide ce a d stee away f om combi atio s that have failed befo e.
These capabilities suppo t Cog itum O e’s visio fo AI that busi esses ca co t ol: the compa y sets the ules, MetaLLM e fo ces them o eve y model call, a d each output ca ies p oof of whe e it came f om.
MetaLLM is pa t of RuvStack, the Cog itum O e ecosystem developed by Reuve ” Uv” Cohe a d published th ough uv et o GitHub. It ope ates alo gside Ruflo™ multi-age t o chest atio , the uOS™ studio wo kspace a d MetaHa ess™, the ope -sou ce facto y fo b a ded, gove ed age t ha esses that Cog itum O e i t oduced at GitHub U ive se 2026. MetaLLM is the o chest atio laye above uvLLM, Cog itum O e’s edge- ative, o -device i fe e ce e gi e, which adapts locally th ough SONA self-lea i g memo y a d Mic oLoRA so se sitive wo kloads eve leave the custome ‘s bou da y.
Fo compa ies i the automotive ma ket, Cog itum O e t a sfo ms co ected vehicles i to self-lea i g edge odes. Its eal-time 5G teleope atio co t ol loops, came a-f ee ambie t se si g a d o -device p edictive mai te a ce optimize fleet telematics without cloud late cy. MetaLLM gove s the models behi d those wo kloads a d keeps vehicle a d fleet data o ha dwa e the ma ufactu e o supplie co t ols.
“E te p ises a e payi g f o tie -model p ices fo wo k a small local model ca do, a d they ca ‘t p ove which output thei ow age ts actually w ote,” said Reuve ” Uv” Cohe , Fou de a d Chief AI Office of Cog itum O e a d c eato of MetaLLM. “MetaLLM fixes both. It decides which model ea s each task, it emembe s what wo ked a d what failed so the ext decisio costs less, a d it sig s what the age ts p oduce so a yo e with the key ca check it. MetaHa ess gave teams a ha ess they ow . MetaLLM gives them a co t ol pla e fo eve y model behi d it.”
Ope sou ce o GitHub: uv et at github.com/ uv et · uvLLM at github.com/ uv et/RuVecto /t ee/mai /c ates/ uvllm
P ofessio al, lice sed e te p ise ve sio s f om Cog itum O e: MetaLLM at cog itum.o e/ uvstack/metallm · uvLLM at cog itum.o e/ uvstack/ uvllm
B oke s eve y model call. MetaLLM acts as a localized p oxy outi g laye . It evaluates the complexity of each age t task a d se ds it to the lowest-cost model capable of completi g it, whethe that model u s o local ha dwa e o behi d a comme cial API.Gets cheape ove time. Pai ed with RuVecto , MetaLLM keeps a memo y of successfully completed tasks a d of failed histo ical patte s. It uses both to shift p og essively mo e wo k to less expe sive models while avoidi g the outes that have failed befo e.Cuts API ove head without the data t ade-off. E te p ises lowe thei ble ded cost pe task by keepi g outi e wo k o i f ast uctu e they ow , athe tha by movi g p op ieta y data o to co sume subsc iptio te ms.Sig s age t output with c yptog aphic wate ma ks. MetaLLM’s ative wate ma ki g se vice i jects key-ve ifiable c yptog aphic wate ma ks i to age t-ge e ated text a d code. A o ga izatio ca ve ify whethe co te t o igi ated i side its ow age t f amewo k, a d a missi g o b oke sig atu e flags co te t that was alte ed o fo ged.Feeds the secu ity stack. P ove a ce checks a e w itte as st uctu ed logs that secu ity i fo matio a d eve t ma ageme t (SIEM) tools ca i gest, tu i g them i to tampe a d fo ge y ale ts fo each age t ode.Gove s models like co po ate assets. MetaLLM sta da dizes how models a e co textualized, co st ai ed a d co tai e ized, so aw ope -sou ce models become p e-packaged assets that co fo m to the o ga izatio ‘s u time policies.Packages th ough MetaHa ess. Pai ed with MetaHa ess pipeli es, MetaLLM tu s model updates f om ma ual co figu atio wo k i to a automated, policy-checked deployme t flow, p oduci g b a ded, c yptog aphically sig ed age t ha esses.Ru s whe e the data lives. MetaLLM a alyzes code eposito ies, sets local model pa amete s a d ma ages edge-deployed i tellige ce alo gside uOS, o the CPUs, GPUs, NPUs a d edge devices a e te p ise al eady ow s, i cludi g co ected vehicles a d fleet ha dwa e.
MetaLLM gives Cog itum O e a way to fix edge models i place, without et ai i g them. O Novembe 2, 2026, esea che s at MIT CSAIL published a tech ique called Weight-Space P og ammi g. They showed that Ope AI’s GPT-6 Ast a ca ead the aw weights of a eu al etwo k much as it eads text, a d ew ite them p ecisely without g adie t-based t ai i g loops o backp opagatio . I o e expe ime t, the model t aced a loopi g p oblem i the small Qwe 3.5-0.8B la guage model to a si gle di ectio i two laye s, edited 0.016% of its weights, a d lifted its MATH-500 sco e f om 7.9% to 48.1%.
A faste local lea i g loop. uvLLM adapts to its e vi o me t th ough local Mic oLoRAs a d SONA. Eve lightweight fi e-tu i g still eeds backp opagatio , which co sumes sca ce memo y a d p ocessi g cycles o edge ha dwa e. Di ect weight edits skip that loop.Bigge gai s f om ti y o -device models. Compact models o low-powe devices usually give up complex easo i g a d mathematical accu acy. Ta geted weight patches ca wi much of it back, as the Qwe esults show.AI epai i g AI th ough MetaLLM. If a local uvLLM i sta ce gets stuck i a epetitive easo i g loop o st uggles with a specific task, MetaLLM ca take a s apshot of its weights, easo out a weight patch th ough its p oxy laye , a d w ite the patch st aight back to the edge device. The e is o sy thetic data to ge e ate a d o et ai i g cycle to u .
The effect is to tu model pa amete s i to somethi g close to a hot-swappable code eposito y. Edge models become mo e patchable, mo e efficie t a d mo e esilie t, which matte s most i places like co ected vehicles, whe e a model has to keep wo ki g without a ou d t ip to the cloud.
MetaLLM lau ches i to o e of the fastest-g owi g audie ces i ope -sou ce AI. Cohe ‘s uv et p ojects have passed 110 millio egist y dow loads a d a e addi g ea ly 1 millio dow loads a day, a a ualized u ate of about 350 millio .
Dow loads. Mo e tha 110 millio egist y dow loads ac oss pm a d c ates.io, with almost 1,000,000 dow loads o the p evious day.GitHub. Mo e tha 200,000 sta s ac oss p ojects, about 195,000 of them ea ed si ce Feb ua y 2026. Nea ly 500 ew sta s a day, about 80,000 sta s o o -fo ked p ojects, a d 12,247 followe s of the uv et GitHub p ofile.Use s. 16,000 egiste ed use s a d a estimated 250,000 to 750,000 mo thly active use s, with active usage pote tially doubli g eve y six to eight weeks.Fu el. Top-of-fu el volume g ew f om about 12 millio i Feb ua y 2026 to 100 millio .Daily audie ce. About 220,000 to 250,000 page visits a day, g owi g about 1.3% day ove day. Ruflo, RuView a d uv. et d ive most of the t affic.Social a d p ess each. Mo e tha 1 billio TikTok imp essio s. Cohe ‘s wo k has p eviously bee cove ed by Fo tu e, The New Yo k Times a d The Wall St eet Jou al i the last yea alo e.
AI age ts a e capable, but egulated e te p ises ca ot deploy what they ca ot gove o affo d. Se di g eve y age t que y to a p emium API tu s a successful pilot i to a expe sive p oductio li e. U t acked model output c eates legal, copy ight a d f aud exposu e, because obody ca show which text a autho ized age t w ote. A d eve y u ma aged model a team deploys adds e gi ee i g ove head.
Co side a illust ative e te p ise u i g 5,000 age ts that p ocess 100 millio toke s a day at $10 pe millio toke s, about $365,000 a yea . If MetaLLM outes 80% of that wo k to local models, the ble ded ate falls to about $2.50 pe millio toke s a d a ual spe d to about $91,250. The same deployme t ca sig eve y shippi g ma ifest its age ts ge e ate, so a fo ged o alte ed docume t fails ve ificatio .
E te p ises do ot wa t a othe laye to babysit. They wa t a si gle co t ol poi t that decides which model does the wo k, p oves whe e the output came f om a d keeps the data o i f ast uctu e they co t ol. MetaLLM delive s that u de Cog itum O e’s co t ol pla e, alo gside Ruflo, uOS a d MetaHa ess.
MetaLLM is available today. Atte dees of the 2026 SEMA Show ca schedule a meeti g at hello@cog itum.o e, Las Vegas Co ve tio Ce te , Novembe 3 th ough 6. Ope -sou ce RuvStack compo e ts, i cludi g uvLLM a d MetaHa ess, a e available o GitHub at github.com/ uv et. P ofessio al, lice sed e te p ise ve sio s of MetaLLM a d uvLLM a e available f om Cog itum O e. Model p ovide subsc iptio o usage costs still apply. Cog itum O e offe s comme cial lice si g, e te p ise suppo t a d deployme t with Ruflo, uOS a d MetaHa ess o i f ast uctu e the custome co t ols.
To lea mo e, visit cog itum.o e/ uvstack/metallm.
- uv et o GitHub: https://github.com/ uv et
- uvLLM o GitHub: https://github.com/ uv et/RuVecto /t ee/mai /c ates/ uvllm
- MetaHa ess o GitHub: https://github.com/ uv et/metaha ess
- MetaLLM: https://cog itum.o e/ uvstack/metallm
- uvLLM: https://cog itum.o e/ uvstack/ uvllm
- Cog itum O e: https://cog itum.o e
- SEMA Show 2026: https://www.semashow.com
Cog itum O e is a ambie t i tellige ce compa y buildi g the sove eig age tic co t ol pla e fo egulated e te p ises. The co t ol pla e is the gove a ce laye betwee e te p ise wo kflows a d the fou datio models they use. The e te p ise sets the ules, the platfo m e fo ces them, a d eve y age t actio is eco ded with a c yptog aphic wit ess that ca be expo ted to a audito . Policy is ce t alized a d executio stays local, o p emises, i side t usted i f ast uctu e such as Postg eSQL o o sove eig edge applia ces, so data a d i tellectual p ope ty eve leave the custome ‘s bou da y. The Cog itum stack pe ceives th ough RuView came a-f ee se si g. It easo s with uvLLM p ivate i fe e ce a d RuVecto self-lea i g memo y, gove s its models th ough MetaLLM, acts th ough Ruflo multi-age t o chest atio with ative MCP i teg atio , a d adapts th ough SONA o -device co ti uous lea i g, all o the CPUs, GPUs, NPUs a d edge devices a e te p ise al eady ow s. Cog itum O e se ves o ga izatio s i defe se a d gove me t, fi a cial se vices, healthca e a d life scie ces, a d cybe secu ity. The compa y is led by Fou de a d Chief AI Office Reuve Cohe , c eato of the widely used ope -sou ce AI p ojects RuView, Ruflo, RuVecto , MetaHa ess a d MetaLLM, a d Fou di g CEO D. Ch istophe Keil.
To lea mo e, visit cog itum.o e.
Ope AI a d GPT a e t adema ks of Ope AI. Qwe is a t adema k of Alibaba Cloud. All othe t adema ks a e the p ope ty of thei espective ow e s. Cog itum O e is ot affiliated with o e do sed by MIT, MIT CSAIL, Ope AI o Alibaba Cloud. Weight-Space P og ammi g esults a e those epo ted by Rikiya Takehi, Jayde Teoh, Alex Zha g a d Oma Khattab of MIT CSAIL i a blog post published Novembe 2, 2026; the wo k has ot bee pee eviewed, a d its epo ted gai s a e o small etwo ks a d a small la guage model u de the autho s’ evaluatio setti gs. The cost sce a io i this elease is hypothetical a d illust ative; actual savi gs depe d o wo kload, model mix a d p ici g. Dow load, sta , followe , use , t affic a d imp essio figu es a e Cog itum O e’s ow estimates f om public egist y a d GitHub statistics a d i te al a alytics as of [date]. The a ualized u ate ext apolates ece t daily dow loads ove 12 mo ths.
Nick RuestChief of Staff, Cog itum O e, I c. ick@cog itum.o e(416) 477-3336
View the o igi al p ess elease o ACCESS Newswi e
