Ɗoƴƴere : Jeonja Ilbo, winnduɗo Lee Won-ji
Ɗoƴƴere timmunde nde winndaa “CLEVI, e model mum waɗɗo from scratch, GAIA 79.07%... hakkunde modelluuji 3,090, o waɗi e hakkunde 2.5% ɓurɗe”
Ñalnde yaltugol : 2026-04-07
Jokkorde : https://www.etnews.com/20260407000203
cip-5.5-agent·cip-5.5-mm stack modelaji keɓtoraaɗi e agentooji 5 fof ɓuri 70
So ɓeydii ko majju, jaabawol moƴƴol 98%+, ɓuri yimɓe (92%)
Startup AI 'CLEVI' waɗi jaabawol moƴƴol 79.07% e benchmark agent AI aduna 'GAIA', huutoraade model mum waɗɗo from scratch, tee o waɗi e hakkunde 2.5% ɓurɗe hakkunde modelluuji 3,090 binndaaɗi fof.
No ɓeyduɓe ngal fannu mbiyi, e suudu benchmark AI, GAIA ena heɓa yettugol sabu o hollata no feewi waawde 'agent AI huutorteeɗo'. Meta AI, HuggingFace, e jaɓɓorgal Paris-Saclay waɗi benchmark ɗo e gootal; o holliraa ko adannde e Noweemburu 2023.
Ko waɗi GAIA seertaa e benchmarkoji goɗɗi ko ko tati. Arannde, sabu jaabawuuji ɗi cuuɗii, waawde overfitting walaa. Ɗiɗaɓere, sabu ɗum naamnii miijagol lesdiiji keewɗi e multimodal, waawde heɓde puccu mawɗo e pattern matching tan walaa. Tatiɓere, e leaderboard Hugging Face resmi, modelluuji 3,090 ɓurɗi e aduna njogora e ɗoƴƴere gootere.
Testset ɗin waɗi e naamne 301 fof. Level 1 naamnii huutoraade kaɓirgol gootol e miijagol laaɓɗol; Level 2 naamnii hawrude kaɓirɗe keewɗe e miijagol hakkunde; Level 3 ko naamne ɓurɗe caɗɗe, ɗe naamnii siryagol e waɗde agent e leƴƴi ɓuri 5. Nde benchmark ɗin holliraa, model GPT (e plugin) heɓi tan ko 15%; jooni, gollotooɓe ɓurɓe njogii ɗum hakkunde 70–80%.
E hakkunde ɗum, CLEVI wi'i ko ɓuri waɗde yaawde e ɗoƴƴere ndee woni: o huutoraali API LLM wuro goɗɗo hay gooto, hono GPT, Claude, walla Gemini. Ko seertinɗum woni huutoraade modelluuji ɗiɗi ɗi o waɗi e hoore mum, e laawol from scratch.
cip-5.5-agent ko ɓeyditorde AI agentic, nde waɗata ko ɓeyditorde hakkille mawnde e pipeline agent ɓeydude ɗaɓɓitannde ɗoɗɗe, waɗde ɗum e hoore mum, e ƴeewtude ɗum. cip-5.5-mm ko model multimodal kuutortooɗo e dow, waawde faamde e miijagol faayiluuji cuɓoraaɗi, hono daande, natal e widewoo. Sababu ko ɓurɗo e tambital GAIA ina waɗi naatnal non-text, hono PDF, natal e faayiluuji daande, waawde multimodal ngal waɗi ko ɓurɗo waɗde ɓeydude njoɓdi.
CLEVI huutori e ɗee modele ɗi tagi e hoore mum ngam naatnude e GAIA e ceertuɗe 5 agent, tee fow heɓi ɓuri 70.
No ɓeyditorde nde holli, e ƴeewtagol caggal waɗangal e nder CLEVI, ko ɓuri hiɗaaki feeñi. E tambital 301 fow, won ɗo daliilu goonga ɗum waawaa heɓde e lowre web jooni. Ko humpito ɗo waawnoo heɓde e internet walla e engine yiytugol ko ittiraa, waylaa, tee waawaa heɓde kadi. So ƴeewtaama ɗum kadi e hakkunde humpito ɗo neɗɗo waawata ƴeewtude e yeeso e jooni, njoɓdi goonga CLEVI waɗi ɓurɗo 98%. Ɗum ko ɓuri njoɓdi hakkunde yimɓe 92%.
Neɗɗo CLEVI wi'i: “CLEVI heɓii 5 agent fow 70+ e benchmark feeñɗo, e naatnude e ɓeyngu 2.5% ɓurɗo yeeso, tan e modele tagiraaɗe e hoore mum from scratch e soliyuushinuuji AI agent mum, tawa alaa cuɓagol modele yaasi. Ɗo yeeso, njoɓdi resmi waawi ɓeydude kadi sabu ƴeewtagol modele e soliyuushinuuji agent tagiraaɗe e hoore mum. Ɗum ko golle mawnde sabu nde ƴeewtii e benchmark feeñɗo e aduna, hollude 'goonga ƴeewtaaɗo'; sabu ko njoɓdi ɓeyditorde AI e Koree tagi e hoore mum, e dow modele tagiraaɗe e hoore mum from scratch; sabu ko sakke modele mum tan, wonaa cuɓagol modele yaasi; tee sabu waawde ɓeydude maanaa njoɓdi ɓurɗo njoɓdi ɗum e yimɓe, nde ƴeewtina ko humpito won e tambital ɗo majjinii.”