Fibea: Etoday, Ɔkyerɛwfo Lee Won-ji
“Clevi, wɔde from scratch wɔn ankasa model ayɛ GAIA 79.07%... wɔ model 3,090 mu ɔwɔ soro 2.5%” Nsɛm no nyinaa a wɔafa aka
Da a wɔyii no adi: 2026-04-07
Link: https://www.etnews.com/20260407000203
cip-5.5-agent·cip-5.5-mm wɔn ankasa model stack, agents 5 no nyinaa anya bɔne 70 anaa nea ɛboro saa
Sɛ yɛhwɛ nsɛm a wɔahwere no a, mmuae a ɛyɛ nokware yɛ 98%+, na ɛboro nnipa (92%) so
AI startup 'Clevi' de from scratch wɔn ankasa model dii dwuma na wonyaa mmuae a ɛyɛ nokware 79.07% wɔ wiase nyinaa AI agent benchmark 'GAIA' mu, na wɔde wɔn sii soro 2.5% mu wɔ model 3,090 a wɔakyerɛw wɔn nyinaa mu.
Sɛnea nnwuma no mufo kyerɛ no, wɔbu GAIA sɛ ɛda 'AI agents a wotumi de di dwuma ankasa' no tumi adi yiye wɔ AI benchmark gua so. Meta AI ne HuggingFace, ne Paris-Saclay Sukuupɔn ne afoforo na wɔyɛɛ benchmark yi bom, na wɔdaa no adi nea edi kan wɔ Ɔpɛpɔn 2023 mu.
Nneɛma atitiriw abiɛsa na ɛma GAIA da nsow fi benchmarks afoforo ho. Nea edi kan, esiane sɛ wɔnkɔda mmuae no adi nti, ɛrentumi nyɛ yiye sɛ wɔbɛyɛ overfitting. Nea ɛto so abien, esiane sɛ ɛhwehwɛ multi-step ne multimodal nsusuwii a ɛyɛ den nti, wontumi nnya nkonta a ɛkorɔn denam pattern matching kɛkɛ so. Nea ɛto so abiɛsa, ɛnam Hugging Face official leaderboard so na model bɛboro 3,090 a ɛwɔ wiase nyinaa si akan wɔ tebea koro mu.
Test set no yɛ nsɛmmisa 301 nyinaa. Level 1 yɛ tool baako a wɔde di dwuma ne nsusuwii a ɛnyɛ den, Level 2 yɛ tools pii a wɔaka abom ne nsusuwii a ɛwɔ mfinimfini, na Level 3 yɛ nsɛmmisa a ɛyɛ den paa a ɛhwehwɛ agent nhyehyɛe ne dwumadi a ɛyɛ anammɔn 5 anaa nea ɛboro saa. Bere a wɔde benchmark no bɔɔ dawuru no, na GPT model (a plugin ahyɛ mu) gyina hɔ ma bɛyɛ 15% pɛ, nanso mprempren akuw a wɔwɔ soro no ama aba 70–80% mu.
Wɔ eyi mu no, Clevi sii so dua sɛ ade a ɛsɛ sɛ wɔhwɛ so paa wɔ mfiridwuma mu wɔ nea wɔanya yi mu ne sɛ wɔannyɛ dwuma biara wɔ external LLM API te sɛ GPT, Claude, ne Gemini so. Ade titiriw ne sɛ Clevi de model abien a wɔyɛɛ no wɔn ankasa denam from scratch kwan so na ɛyɛɛ adwuma.
cip-5.5-agent yɛ AI nhwɛsoɔ a ɛyɛ adwuma sɛ ɔhwɛfoɔ, na ɛyɛ adwuma sɛ adwene titiriw wɔ ɔhwɛfoɔ nhyehyɛeɛ mu a ɛtumi yɛ nnwuma a ɛyɛ den ho nhyehyɛeɛ (planning), yɛ (execution) ne nhwehwɛmu (verification) no ankasa. cip-5.5-mm yɛ multimodal nhwɛsoɔ a ɛwɔ tumi kɛseɛ na ɛte fael ahodoɔ te sɛ nne, mfonini ne video ase na ɛyɛ ho adwene. Esiane sɛ GAIA nsɛmmisa no mu pii ka PDF, mfonini, ɔdio fael ne nneɛma foforɔ a ɛnyɛ nsɛm nkutoo ho nti, saa tumi a ɛfa nneɛma ahodoɔ ho no bɛyɛɛ ade titiriw a ɛmaa wonyaa nkonimdie kɛseɛ.
Klevi de saa nhwɛsoɔ abien a ɛyɛ wɔn ankasa yi gyinaa so de ɔhwɛfoɔ nhyehyɛeɛ ahodoɔ anum a ɛsono wɔn ho wɔn ho kɔdii GAIA akansi no, na wɔn nyinaa nyaa bɔɔ 70 anaa nea ɛboro saa.
Sɛnea adwumakuw no kaeɛ no, Klevi mu akyirehwɛ maa wohuu nea ɛyɛ anigyeɛ. Nsɛmmisa 301 no mu bi, mprempren wɔrentumi nhu adanseɛ a ɛkyerɛ mmuaeɛ a ɛyɛ nokware wɔ wɛbsaet a ɔmanfoɔ bɛtumi anya so no so. Ɛyɛ tebea a wɔapopa anaa wɔasesa nsɛm a na wɔtumi nya wɔ intanɛt anaa mfiri a wɔde hwehwɛ nsɛm mu kane no, enti mprempren wɔrentumi nnya bi bio. Bere a wɔsan susuw ho wɔ nsɛm a nnipa tumi hwɛ wɔ ɔmanfoɔ anim mprempren no mu no, Klevi mmuaeɛ a ɛyɛ nokware no bɛyɛɛ ɔha mu nkyem 98 anaa nea ɛboro saa. Eyi yɛ dodow a ɛboro nnipa wɔ ɔmanfoɔ mu no mu nkyem 92 mprempren.
Klevi ɔman no kae sɛ, “Klevi nnyaa nkabom biara amfa abɔnten model ahodoɔ; mmom, yɛde yɛn ankasa model a yɛyɛɛ no from scratch ne yɛn ankasa AI agent ano adwuma nkutoo yɛɛ adwuma, na yɛn agents anum no nyinaa nyaa 70+ de kɔɔ ɔmanfoɔ benchmark no mu ɔha mu nkyem 2.5 a ɛwɔ soro no mu. Ɛda adi sɛ daakye, sɛ yɛkɔ so ma yɛn ankasa model ne agent ano adwuma no tu mpɔn a, yɛbɛtumi ama aban score no nso akɔ soro bio. Saa nkonimdie yi ho hia paa efisɛ wɔsusuwii wɔ wiase nyinaa ɔmanfoɔ benchmark mu de kyerɛɛ ‘ahotosoɔ a wɔagye ato mu’; ɛyɛ nkonimdie a AI adwumakuw a ɛwɔ ɔman yi mu anya denam wɔn ankasa model a wɔyɛɛ no from scratch so; ɛyɛ nea ɛfiri model stack a ɛyɛ wɔn ankasa, na ɛnyɛ abɔnten model ahodoɔ a wɔaka abom; na esiane nsɛm a ayera wɔ nsɛmmisa no mu bi nti, ɛwɔ nkyerɛaseɛ mu mfasoɔ a ɛboro score no so.”