From-scratch Foundation Model klébi AI(ivy)

“Hol no cuɓoraaɗo tokko waawi waɗde model hunuko AI mawɗo ɓurɗo heɓde daraja e aduna?”

No waɗiraa model hunuko mawɗo e startup pamarel?

Ɗum woni naamnal ɓurɗo heɓde caggal nde klébi hollitii model Foundation jogiiɗo 1.4 trillion paramitaaji.

Ko ɓeynguɗe ɗe waɗata tan fine-tuning e modelaji open-source, klébi waɗii ɗum fof e hoore mum, daga collecte données haa e suɓaade model, jannginde e distributed scale mawɗo, e ƴeewtaade kalite.

En hollan e les ɗoo sirru ɗum e semmbe mum no feewi.

1. Startup pamarel waawi waɗde model hunuko mawɗo?

E lewru Juyye 2025, nde klébi udditii e ahịa model hunuko mawɗo waɗaaɗo e hoore mum, yimɓe heewɓe (huutorɓe) ɓe waɗii hayre e sikki. “Ko ɗum waɗaaɗo e hoore mum goonga?” “Ko o waylii model open-source seɗɗa tan?”

E goonga, kompaniiji heewɗi e leydi e ɗoɓɓere ɓe waɗata tan transfer learning (fine-tuning) e model open-source, ɓe waɗa ɗum no model mum'en.

Amma klébi hollitii **‘From-scratch Foundation Model’**.

Ɗum firti ko collecte données, suɓaade model, jannginde e distributed scale mawɗo, e ƴeewtaade kalite, fof ko ɓe suɓii e waɗi e hoore mum'en.

2. Ko waɗi waɗde model hunuko mawɗo saɗi

Ngam jannginde From-Scratch Foundation Model, wajibi ko janngude pelle ɗee fof.

Dataset mawɗo jogiiɗo kalite ɓurɗo moƴƴude ina waɗi

• Diverse: heɓde ɗoɗɗugol e ɗemɗe, domenaji, e nder (binndol, natal, ekn.)
• Ɗeƴƴinde: momtude noise, reende kalite, ƴeewtaade e filteraade keɓe bonɗe walla ɗe jogiiɗe njamndi
• Lisansa: tabbitinde hakkunde binndol e jamiroore huutoraade keɓe

Infrastruktuure computing mawɗe

• Cluster GPU/TPU jogiiɗo semmbe mawɗo: haɗde kaɓirɗe nodeaji heewɗe (ko haa e kumaaji) ngam wallude jannginde e distributed scale e keɓe mawɗe
• Framework jannginde e distributed jogiiɗo yoɓde: DeepSpeed, Megatron-LM, Ray e goɗɗe
• Storage/network: naatnude e yaltinde keɓe mawɗe, e environment network yaawɗo

Suɓaade architecture model

• Jëfandikoo cosaan yu bees: Transformer, MoE(Mixture of Experts), multimodal, ak leneen
• Scalability: nattal baaxalug limub parameters, layer, juura input, ak leneen
• Efisiyaas: gaawug jàng ak xam, ak sàmm mémoire

Pexe yi ak algorithmes yu jàng

• Tàggatug pretraining: language model (misal: next token prediction), multimodal (misal: contrastive learning), ak leneen
• Fom yu optimization: mixed precision, gradient accumulation, learning rate schedule, ak leneen
• Normalisation/jàppale: dropout, layer norm, weight decay, ak leneen

Sistemub natt ak seetlu

• Setu benchmarks: jëfandikoo datasets yu standard (Benchmarks)
• Nattu biir: nattu sukkandiku ci scenarii yu ñuy jëfandikoo dëgg-dëgg
• Seetlu bu sax: seet qualité, biais ak kaarange ci jamono jàng bi ak gannaawam

Dooleg nit ñi ak kurél gi

• Jàngat AI/ML: defar model ak sos algorithmes
• Ingeniër yu données: dajale, sellal ak saytu données
• Ingeniër yu infrastructure: sistem yu distribué, saytu cloud/on-premises
• Manageru projet: saytu waxtu, budget ak qualité

Xalaat yu aju ci etik, yoon ak kaarange

• Etik ci AI: biais, kaarange, leerleeral ak verantwoordelijkheid
• Dëppoo ak yoon: sàmm xibaaru bopp, copyright ak suñu moomel données
• Kaarange: aar ci génnug données/model ak kontrolug dugg

Fi nekk, limub waajur yi di liggéey ci xam-xam AI ci àdduna méngoo na ak 2 000 nit, te ëpp ci ñoom dañuy bokk ci kompanii yu mag yi mel ni META, Google ak XAI. Ci réew mi, koox bu mat sëkk buy man a defar model Foundation, dale ko ci dajale données ba ci infrastructure bu mag bu ñuy jàngal, dafa néew lool.

3. Defar modelu làkk bu mag lool ak nit ñu néew.

Ci kaw njàngam ci deep learning ak machine learning lu ëpp 10 at, njiitu I환호 Lee sos na kompanii bi (CLEVI) 3 at ci kanam, ak tënkug “defar AI bi gën ci àdduna”

Mu defar boppam bépp yoon wu aju ci tabaxug pipeline bu données bu kompanii bi, defarug infrastructure deep learning bu distribué bu mag, sosug architecture model ak seetlu qualité, ba am doole yi aju ci defarug Foundation Model.

Gannaaw bi mu jàngale model bi ay yoon yu bari, léegi am na modelu Clevi-5-x bu man a daje ak model yi gën ci àdduna.

Ngam jannginde limɗo konngol mawɗo no feewi, ena sokki infrastrukture nde GPU ɗe ɗoƴƴi e dow kaaɗe ɗe keewi haa ɗuɗɗi, tawa ɗe njokkondiraa e klaster, ngam waawde waɗde gollalaji ɗi heewi haa ley sekondoo e sahaa gooto. E nder ɗum, teknoloojiiji ɗi njogii daraja mawɗo ko ɗiɗoƴƴinde gollalaji e nokkuure keewnde e ɗaɓɓude ɗoƴƴugol jokkondiral. Haa jooni, e nder lesdi ndee, himo waɗi waawde jannginde limɗe ɗe no feewi sabu alaa “algorithme ɗoƴƴugol jogorɗe”. Ardo Clevi, Lee Hwan-ho, waɗi algorithme jogorɗe gollalaji ɗi njokkondiraa e dow kisal mum, tee o waɗi e hoore mum jannginde limɗo konngol mawɗo ɗo heɓi 1.4 trilliyɔn (1.4 Trillion) parametere, ko adannde e lesdi ndee.

4. Sababu nde waawi waɗde limɗe keewɗe e yimɓe gollotooɓe seɗɗa

Kompanii big tech ɗuɗɗi kadi njogii teknoloojiiji jogorɗe infrastrukture mawnde e laɓɓinde keɓe.

Ngam jogorde ɗum, ena sokki heɓde ñaawoore nde jogii limɗoɓe jannginɓe ɗoƴƴi hakkunde ɗeeri e jiɗi.

Ammaa, Clevi ena waɗa tee ena jogii limɗe keewɗe, tawa o waɗata ɗum e yimɓe jannginɓe seɗɗa.

No ɗum waawi waɗde?

Clevi ena waɗa tee ena jogii limɗe keewɗe huutoraade teknolooji Teacher-Student Model(Knowledge Distilation), ngam waɗde limɗe keewɗe e yimɓe gollotooɓe seɗɗa.

Ndeen, en ƴeewto teknolooji Teacher-Student.

Teacher-Student(laawol feccugol anndal)

Teknolooji Teacher-Student ko e laawol weeɓi, limɗo mawɗo (Teacher Model) ena ƴeewta tee ena hokku jaŋtorde limɗo ɓiɗɗo (Student Model), ngam waawde waasde yimɓe jannginɓe ɗoƴƴi ɗo heewi e yimɓe seɗɗa, tee waɗde limɗe keewɗe e yaawde.

• Mother Model(limɗo mawɗo) ena ƴeewta tee ena hokku jaŋtorde limɗe ɗe njogii ko feere-feere, tee ɗum waasda yimɓe jannginɓe ɗo heewi.
• So en jannginii e laawol ngal huutoraade Clevi-x-platform, en waawi jannginde e waasde limɗe jogii semmbe, hono Reasoning, VLM, Physical AI, Coding Agent, e nder balɗe 10 haa 15.

Clevi Coding Agent

Clevi Phsycal AI Learning Platform

VLM-Vision Language Model

Agent kisal

5. Feere nde Clevi-x-platform waɗi e teknolooji e ƴeewtagol kalite

Semmbe limɗo e waawde ɓeydude

• Ɓurɗe Ɗaɓɓugol (CoT/Reasoning): cip-5-x waɗata laawol cuɓoraaɗo ngam waɗde ɗaɓɓugol e waɗde dalillaaji e darnde kuutorgol, tee hollata waawde waɗde e fassitde kalkulasyoon ɗo hoolaare woodi e caɗeele saayansi, matematiki e injiniiriya. E dow benchmarkuuji keessaa amen, ɗum waɗetee e ɗaɓɓitaare ngam heɓde semmbere ɓurɗe GPT-5 walla ɗo fota e ɗum, tee ƴeewtagol e waawde ƴeewtaade keɓe e ɗoɗɗugol prompt gooto ko jeyaaɗe e cuɓoraaɗe kalite mawɗe ɗe njogii.

• Fulleere multimodal timmunde: VLM jaɓɓugol e yahrude e huutoraade (cip-5-vision), model multimodal mawɗo ngam ɗaɓɓugol keɓe (ivy-4-mm), e model ɗo ɓuri newnude ngam kuutorgol e ɗowrowol (ivy-3-text) mbaawi waɗde e jokkondirde e platform gooto, ngam suɓaade jokkondiral ɓurɗo moƴƴude e dow no gollal ngal waɗiraa (yiytude/renndinde/taƴtinde vs. ɗaɓɓugol/fassitde/huutoraade).

Waawde huutoraade e enterpraayzi

• Kisal e waawde heɓde e on-premises: huutoraade e cuɓoraaɗe kisal ɗo geese ngalaa (naatnugol-jannginde-golle-ndaarndugol), cuɓoraaɗe HA, ɓeydude e moduluuji, e SVCE (environment kisal kuutorgol ceertuɗo) ngam seertinde e kisnude keɓe e parameetere mudel.
• Naatnugol e ɓeydude yaawde: Ivy Chat (jokkondiral multimodal e agent ngam gollal), API Platform (SaaS e cuɓoraaɗe aduna) mbaɗata naatnugol e huutoraade yaawde.

Ɓurɗe Physical AI

• Jokkondiral simulation ɗo NVIDIA Isaac e reinforcement learning ɗo ƴeewtugol (Evolutionary RL), e jannginde ɗoɗɗude e Sim2Real e jannginde keɓe sintine e sahaa gooto, ɓeydii e waawde janngude e jaɓde e nokkuure. Waɗde ɗum waɗa digital, janngude robotik e neldugol e gollal fow ko ko ɓuri heɓde teddungal e hakkunde cuɓoraaɗe.

Kisal gollal e laawol njogoraagu

• Ɗaɓɓugol sariya mudel/prompt/kaɓirɗe, evaluation suite (anndal e Koreere e Angaleere, gollal ɗoɗɗe e industries, keɓe juuyre e kisal), njogoraagu bayle (SLO/SLA), e ƴeewtagol gollal (latency, limgal jamma, TCO) ɗaɓɓittee e laawol platform gooto.

Jooni, e Koree, ena woodi ɓurɗo 300 e gollotooɓe AI ɗo njogii gollal mudel, kono

CLEVI tan woni ɗo waawi hokkude e sahaa gooto gollal on-premises e cloud, e mudel foundation mum waɗaa e semmbere mawnde.

6. No huutoraade mudel ɗemngal e industries

Naatnugol AI ko waɗde cuɓoraaɗe mawɗe e ƴeewtagol laɓɓungal, ɗum waɗi ko foti waɗde ƴeewtagol e ƴeewtaade e hoore maa ngam anndude so solution ɗum wallata gollal enterpraayzi e goonga.

• CLEVI wona tan waɗa modele, kon o hokku solushon AI ɗe waawi huutoraade e gese golle goonga.
• Misal, o hebbii kanaliiji enterpraayz timmuɗi, ɗi waɗi Ivy Chat Platform, Clevi API Platform, keɓtinde ɗe cuɓaaɗe ngam jeyaaɗe e ngànhuuji, e jaabawol goonga ngam ɗaɓɓitanɗe kisal.
• O heɓi gandal teknolooji ngal yiɗetee waawde heɓaade e ley e lesdiiji goɗɗi: AI waɗɗo e aduna, robotik, e huutoraade keɓe ɗe ɗaɓɓitii geɗe keewɗe.

CLEVI hokku solushon AI goonga, ɗe AI woni ‘laawol’ tan, wonaa ‘waɗde’, ngam wallude e ɓeydude semmbe golle e ɓeydude kesɗitinal ngànhuuji. Ɓeydude e tawaade cuɓoraaɗe From-scratch Foundation model goonga, tawa ɗum e hoore mon.

Ɗaɓɓitannde e ɗaɓɓugol demo: [email protected]

Copyright© 2025 Clevi Inc. All rights reserved.