IHuawei imemezele namuhla ukuthi imodeli ye-Pangu openPangu-2.0-Pro enamapharamitha angu-5050 billion, okuhlanganisa izisindo zemodeli nekhodi eyisisekelo yokwenza iziphetho, kanye nombiko wezobuchwepheshe, sekuvulwe umthombo kwethulwa ngokusemthethweni.
I-openPangu, imodeli ye-AI evulwe umthombo yakwaHuawei, iwuphawu lwamamodeli e-AI akwaHuawei oluzinikele ekunikezeni imboni izindlela ezihamba phambili zokusebenzisa i昇腾 ngobuchwepheshe bokuqeqesha nokwenza iziphetho obudalelwe i昇腾.
I-openPangu-2.0-Pro iyimodeli yolimi enkulu ye-mixture of experts (MoE) eqeqeshwe kusetshenziswa i昇腾 NPU, enesilinganiso samapharamitha esingaba ngu-505B, kanti amapharamitha asebenzayo ku-token ngayinye angaba ngu-18B. Imodeli isekela ubude bomongo obungu-512k, futhi inani ledatha yokuqeqesha lilonke lingaba ngu-34T tokens. Esigabeni sokuqeqesha kwangemva kwalokho kuqedwe ukulungiswa okuhlanganisa imodi esheshayo nevilaphayo (SFT), ukuqiniswa kokufunda ngezihloko eziningi (RL), kwase kuqedwa ukuhlanganiswa kwamakhono nge-online distillation (OPD).
I-openPangu-2.0-Pro yenze ukuthuthukiswa okuphelele ekwakhiweni kwemodeli:
Ukwakhiwa kwe-Attention: iqhubeka nokusebenzisa i-MLA esebenza kahle, futhi isebenzisa ukwakhiwa okuxubile okuzimele, okuhlukaniswe ngamaleveli, kwe-DSA+SWA, ngesilinganiso samaleveli esingu-1:2; ungqimba lwe-SWA lubhekele ukumodela iwindi lendawo, kanti ungqimba lwe-DSA lubhekele ukuhlanganisa umhlaba wonke okungasindi. Lokhu kwehlisa kakhulu izindleko zokubala, imemori yevidiyo nokufinyelela kwimemori ekwenzeni iziphetho ngokulandelana okude, kuyilapho kugcinwa ukunemba.
Ukwakhiwa kwe-topology: ukuxhumana kwendabuko kwe-residual kuthuthukiswe kwaba ukwakhiwa kwe-mHC okunemifudlana emi-4, okuthuthukisa ukuhlukahluka kokumelela namandla okwenza kahle jikelele.
Imojula ye-self-speculative: isebenzisa ukwakhiwa kwe-MTP enamakhanda angu-3, ibikezela ama-token angu-3 engeziwe ngesikhathi esisodwa, okwandisa kakhulu isivinini sokwenza iziphetho kwemodeli.
I-Muon optimizer: ngesikhathi sokuqeqesha kusetshenziswa i-Muon optimizer, okuthola isivinini sokusondela ngokushesha.
Enkulumweni eyinhloko yomhlangano we-Huawei Developer Conference HDC 2026 ngoJuni walo nyaka, u-Yu Chengdong, ongumqondisi ophethe wakwaHuawei, usihlalo we-Product Investment Review Committee kanye nosihlalo we-Terminal BG, wamemezela ukuthi uhlelo lwe-openPangu 2.0 luzovula izingxenye ezinkulu ezingu-7 ngokulandelana kusukela mhla zingama-30 kuJuni, okuhlanganisa ikhodi yokuqeqesha kwangaphambili, ikhodi yokuqeqesha kwangemva kwalokho nama-operator okuqeqesha avulwa okokuqala.
Ngomhla zingama-30 kuJuni, imodeli ye-openPangu-2.0-Flash enamapharamitha angu-92B yavulwa umthombo futhi yethulwa ngokusemthethweni.
Mayelana nenani eliphelele lamapharamitha angu-505B le-openPangu-2.0-Pro kuphela, u-Yu Chengdong wake wachaza ukuthi amandla okubala amaningi asekele izidingo zezinye izinkampani zasekhaya, ngakho inani aligcinelwe bona lilinganiselwe kakhulu; ngaphezu kwalokho, izindleko zamandla okubala e-AI ziphezulu kakhulu, futhi uHuawei ugxile kakhulu ekuthuthukiseni ukubambezeleka nokusebenza kokudlulisa.
Izingxenye ezihlobene nemodeli ye-openPangu-2.0 zilayishwa ngokulandelana ezinkundleni ezivula umthombo. Ikheli lomthombo ovulekile:
https://ai.gitcode.com/ascend-tribe/openPangu-2.0-Pro