Imodeli eyisisekelo yobuhlakani bokwenziwa enkulu kunazo zonke eSouth Korea: I-LG AI Research Institute ikhiphe i-K-EXAONE 2.0 enamapharamitha angu-750B

I-LG AI Research Institute ikhiphe imodeli ye-K-EXAONE 2.0 ku-Hugging Face ngesikhathi sendawo namuhla. Njengoba igama layo libonisa, lena imodeli yesibili eyakhiwe yi-LG AI Research Institute ngaphansi kohlelo lwamamodeli esisekelo obuhlakani bokwenziwa obuzimele loMnyango Wezesayensi Nobuchwepheshe Bolwazi Nokuxhumana waseSouth Korea.

I-K-EXAONE 2.0 isebenzisa ukwakheka kwe-MoE okuxubile kokunaka, inamapharamitha angu-750B esewonke kanye namapharamitha angu-37B asebenzayo. Ubukhulu bayo bungaphezu kokuphindwe ka-3 kobemodeli yokuqala i-K-EXAONE engu-236B-A23B, futhi iyimodeli eyisisekelo yobuhlakani bokwenziwa enkulu kunazo zonke eyake yakhiwa eSouth Korea. Imodeli entsha ikhishwe ngomthombo ovulekile ngaphansi kwelayisensi ye-Apache 2.0.

Uma kuqhathaniswa ngokulandelana, i-K-EXAONE 2.0 ithole isilinganiso samaphuzu angu-70.1 ezivivinyweni eziyisisekelo ezingu-24, okungaphezu kokunyuka ngo-10% uma kuqhathaniswa namaphuzu angu-63.3 e-K-EXAONE yokuqala. Ezivivinyweni ezintathu eziyinhloko zokubhala amakhodi nokubhala amakhodi kwama-agent, ukusebenza kuthuthuke cishe ngo-30%.

Uma iqhathaniswa namanye amamodeli, le modeli idlula i-GLM-5.1 ekuqondeni umongo omude, idlule i-GLM-5.1 ye-Zhipu kanye ne-Qwen3.5 ekusebenziseni amathuluzi ama-agent, futhi ekulinganiseni ukulandela imiyalelo ibonise ukusebenza okulingana nokwe-GLM-5.1, DeepSeek V4 Pro Max, Qwen3.5 namanye.