Zyphra Ibuyekezwa: Amamodeli e-MoE Asebenza Ngempumelelo Aqeqeshwe ku-AMD Enza Okungaphezu Kwalokho Okulindelwe Ngobukhulu Bawo

#Quasa #QUA #zyphra

I-Zyphra (evezwe ku-Quasa.io/projects/zyphra) yinkampani ethuthukisa ubuhlakani obudlula obabantu obuvulekile, esebenza kuzo zonke izingqimba zobuchwepheshe futhi eyakha kusukela ekuqaleni izinhlelo ze-AI ezizimele, ezisobala nezihambisana nezimiso zabantu.

Esikhundleni sokuthembela kumamodeli avaliwe nakwihadiwe yomhlinzeki oyedwa, i-Zyphra ithuthukisa amamodeli ayisisekelo avulekile anobuchwepheshe obusha (umndeni we-ZAYA) anobuhlakani obuningi ngokuphawulekayo uma kuqhathaniswa nobukhulu bawo, bese iwanikeza nge-Zyphra Cloud — inika izinhlangano ukulawula kwangempela ingqalasizinda yazo ye-AI.

Emnyombweni wayo, i-Zyphra iqhubekisela phambili imingcele yokusebenza ngempumelelo kobuhlakani: yakha amamodeli anikeza ikhono lokucabanga elisezingeni lamamodeli aphambili kuyilapho esebenzisa amapharamitha asebenzayo ambalwa kakhulu futhi esebenza ngempumelelo kwamanye ama-chip.

Okugqamile okuyinhloko kuhlanganisa:
• ZAYA1-8B — Imodeli ye-Mixture-of-Experts enamapharamitha asebenzayo angu-760M kuphela (angu-8.4B esewonke) eyenza okungaphezulu kakhulu kwalokho okulindelwe ngobukhulu bayo ezivivinyweni zokuqhathanisa amakhono okucabanga okuyinkimbinkimbi, ezibalo nokubhala ikhodi — ngokuvamile ilingana namamodeli amakhulu kakhulu anezisindo ezivulekile noma iwahlule.
• Isakhiwo Esisha se-MoE++ — I-Compressed Convolutional Attention, ukuqondisa okuthuthukile, nobuchwepheshe obusha ekuqeqesheni kwangaphambili nasekuqeqesheni okulandelayo okukhulisa ukusebenza nge-FLOP ngayinye kuze kufinyelele ezingeni eliphezulu kakhulu.
• Ukuqeqesha Okubeka i-AMD Phambili — Amamodeli okuqala amakhulu e-MoE aqeqeshwe kusengaphambili, aqeqeshwa phakathi nenqubo, futhi acwengisiswa kusukela ekuqaleni kuze kube sekugcineni kuma-GPU e-AMD Instinct MI300X nakubuchwepheshe benethiwekhi be-Pensando — okufakazela ukuthi ingqalasizinda ye-AI ezimele, esebenzisa abahlinzeki abaningi, isilungele ukusetshenziswa.
• Zyphra Cloud — Inkundla yazo zonke izingqimba zobuchwepheshe elungele ukusetshenziswa emsebenzini wangempela, eletha ubuchwepheshe obusha be-Zyphra Research kubathuthukisi, ezinkampanini ezinkulu nakubahlinzeki bezinsizakalo zamafu abasebenza ngezinga elikhulu kakhulu, ngokufakwa ekusebenzeni nokwandiswa okushelelayo.
• Amamodeli Aphatha Izinhlobo Eziningi Zokwaziswa Nawesikhathi Esizayo — I-ZAYA-VL (ukubona nolimi), izinguqulo zokuhlola ze-ZAYA1-74B, ucwaningo lwamamodeli olimi asebenzisa indlela ye-diffusion, nokunye ohlelweni lwentuthuko.

Ifaneleka kahle kubacwaningi be-AI, izinkampani ezinkulu ezifuna i-AI ezimele, abathuthukisi abakha besebenzisa amamodeli avulekile, nanoma ubani okhathele ukuboshelwa kumhlinzeki oyedwa noma izindleko ezinkulu zokusebenzisa amamodeli ukuze akhiqize izimpendulo. Ngo-2026, i-Zyphra igqama njengenye yemizamo ethembisa kakhulu yobuhlakani obudlula obabantu obusebenza ngempumelelo, obuvulekile nobungancikile ohlotsheni oluthile lwehadiwe.

Umphakathi uyayithanda:
“I-ZAYA1-8B iyamangalisa — inamapharamitha asebenzayo angaphansi kuka-1B kodwa yenza kahle kakhulu ezivivinyweni zokuqhathanisa amakhono ezibalo nokubhala ikhodi ezihlula amamodeli anobukhulu obungu-10x obayo. Leli yikusasa lobuhlakani obusebenza ngempumelelo.”
“Ekugcineni, sekukhona enye indlela yangempela esikhundleni sokuthembela ku-NVIDIA kuphela. Amamodeli e-Zyphra aqeqeshwe ku-AMD afakazela ukuthi i-AI ezimele ayigcini nje ngokuba yinto engenzeka — isikhona.”
“I-Zyphra Cloud + amamodeli e-ZAYA = iqoqo eliphelele lobuchwepheshe bobuhlakani obudlula obabantu obuvulekile ebesililindile.”

Igqama kakhulu ngobuhlakani obuningi uma kuqhathaniswa nobukhulu bayo, emisebenzini edinga ukucabanga okujulile, ekukhiqizeni izimpendulo ngezindleko eziphansi, nasekwakheni izinhlelo ze-AI ezizimele ngempela ngaphandle kokunciphisa ukusebenza kwazo.

Okushiyekayo: Amamodeli asesezigabeni zokuqala zentuthuko (izinguqulo ezintsha njenge-ZAYA1 zisengezokuhlola); imiphumela engcono kakhulu ivame ukudinga ukucubungula okwengeziwe ngesikhathi sokuhlolwa ukuze kufinyelelwe ukusebenza okuphezulu kakhulu; ukusetshenziswa kwenkundla yamafu ephelele ngezinga lezinkampani ezinkulu kuyakhula.

Sekukonke, ezinhlanganweni nakubathuthukisi abafuna i-AI esezingeni eliphambili evulekile, esebenza ngempumelelo nezimele ngo-2026, i-Zyphra ingenye yezinkampani ezibaluleke kakhulu nezithembisa kakhulu ezisebenza kuzo zonke izingqimba zobuchwepheshe kulo mkhakha. Ayigcini nje ngokuba ngenye ilabhorethri yamamodeli — yakha ingqalasizinda ephelele yesikhathi esizayo yobuhlakani obudlula obabantu obuvulekile.

Zuza nomklomelo we-QUA nge-Quasa!  

Izinkanyezi ezingu-4.8/5 (ivelele ngobuchwepheshe obusha, ukusebenza ngempumelelo nokugxila ekuzimeleni; kunamaphuzu amancane aphathelene nokungakavuthwa ngokuphelele kwamamodeli amasha).

Qala: https://quasa.io/projects/zyphra