Zyphra Ihlolwa: Amamodeli e-MoE Asebenza Ngobungcwethi Aqeqetshwe ku-AMD Enza Okudlula Ubukhulu Bawo

#Quasa #QUA #zyphra

I-Zyphra (evezwe ku-Quasa.io/projects/zyphra) yinkampani eyakha ubuhlakani obuvulekileyo obudlula obabantu, esebenza kuzo zonke izingxenye zobuchwepheshe, isakha amasistimu e-AI azimeleyo, asobala njalo ahambelana lezimiso zabantu kusukela esisekelweni.

Esikhundleni sokuthembela kumamodeli avaliweyo lakumitshina evela kumthengisi oyedwa, i-Zyphra yakha amamodeli esisekelo avulekileyo alobutsha (imuli ye-ZAYA) alobuhlakani obunengi ngokumangalisayo nxa kuqathaniswa lobukhulu bawo, ibisiwenza atholakale nge-Zyphra Cloud — inika izinhlanganiso amandla eqiniso okulawula ingqalasizinda yazo ye-AI.

Okuyikho okuyisisekelo se-Zyphra yikuqhubekisela phambili imingcele yokusebenzisa ubuhlakani ngobungcwethi: ukwakha amamodeli acabanga ngezinga eliphambili kakhulu esebenzisa amaparamitha asebenzayo amalutshwana kakhulu njalo esebenza ngobungcwethi kwamanye ama-chip e-silicon.

Okuqakathekileyo kugoqela:
• ZAYA1-8B — Imodeli ye-Mixture-of-Experts elamaparamitha asebenzayo angu-760M kuphela (angu-8.4B esewonke) eyenza okudlula khatshana ubukhulu bayo ezilinganisweni zokuhlola ukucabanga okunzima, izibalo, lokubhala ikhodi — ivame ukulingana kumbe ukwedlula amamodeli amakhulu kakhulu alezisindo ezivulekileyo.
• Ukwakheka Okutsha kwe-MoE++ — I-Compressed Convolutional Attention, ukuqondisa okuthuthukileyo, lendlela ezintsha zokuqeqetsha kokuqala/lokuqeqetsha okulandelayo ezenza ukusebenza nge-FLOP ngayinye kube sezingeni eliphezulu kakhulu.
• Ukuqeqetsha Okuqalisa nge-AMD — Amamodeli okuqala amakhulu e-MoE aqeqetshwe okokuqala, aqeqetshwa phakathi, njalo alungiswa ngokugcweleyo kusukela ekuqaleni kuze kube sekucineni kuma-GPU e-AMD Instinct MI300X lakubuchwepheshe bokuxhumanisa be-Pensando — efakazela ukuthi ingqalasizinda ye-AI ezimeleyo, esebenzisa imikhiqizo yabathengisi abehlukeneyo, isilungile.
• Zyphra Cloud — Inkundla yokusetshenziswa emsebenzini wangempela ehlanganisa zonke izingxenye zobuchwepheshe, eletha okutsha kwe-Zyphra Research kubathuthukisi, amabhizimusi, labahlinzeki bezinsizakalo zekhompiyutha ezinkulu kakhulu, ilokufakwa ekusebenzeni lokwandiswa okungelazihibe.
• Amamodeli Asebenzisa Izinhlobo Ezitshiyeneyo Zolwazi Lawesikhathi Esizayo — I-ZAYA-VL (ukubona lolimi), amamodeli okuhlola e-ZAYA1-74B, izilingo zamamodeli olimi asebenzisa i-diffusion, lokunye okunengi okusohlelweni lwentuthuko.

Ifanele kakhulu abacwaningi be-AI, amabhizimusi adinga i-AI ezimeleyo, abathuthukisi abakha besebenzisa amamodeli avulekileyo, laye wonke osediniwe yikubotshelwa kumthengisi oyedwa kumbe yizindleko ezinkulu zokusebenzisa amamodeli ukukhipha izimpendulo. Ngo-2026, i-Zyphra ivelele njengenye yezindlela ezithembisa kakhulu kubuhlakani obudlula obabantu obusebenza ngobungcwethi, obuvulekileyo njalo obungabotshelwanga kuhlobo oluthile lwemitshina.

Umphakathi uyayithanda:
“I-ZAYA1-8B iyamangalisa — ilamaparamitha asebenzayo angaphansi kuka-1B kodwa idlula kakhulu ezilinganisweni zokuhlola izibalo lokubhala ikhodi ezihlupha amamodeli alobukhulu obuphindwe ka-10x kulayo. Leli likusasa lobuhlakani obusebenza ngobungcwethi.”
“Ekucineni, sekulenye indlela yeqiniso yokuphuma emhlabeni osebenzisa i-NVIDIA kuphela. Amamodeli e-Zyphra aqeqetshwe ku-AMD afakazela ukuthi i-AI ezimeleyo kayisiyo nto engenzeka kuphela — isikhona.”
“Zyphra Cloud + amamodeli e-ZAYA = iqoqo elipheleleyo lobuchwepheshe bobuhlakani obuvulekileyo obudlula obabantu ebesililindele.”

Ivelela ikakhulu ekubeni lobuhlakani obunengi nxa kuqathaniswa lobukhulu bayo, emisebenzini edinga ukucabanga okunengi, ekukhipheni izimpendulo ngezindleko eziphansi, lasekwakheni amasistimu e-AI azimeleyo ngeqiniso kungancitshiswa izinga lokusebenza.

Ubuthakathaka: Amamodeli asesebangeni lokuqala lentuthuko (asanda kukhitshwa njenge-ZAYA1 ngawokuhlola); imiphumela emihle kakhulu ivame ukudinga ukubala okwengezelelweyo ngesikhathi sokuhlolwa ukuze kufinyelelwe ezingeni eliphezulu kakhulu lokusebenza; ukwamukelwa kwenkundla yonke yezinsizakalo zekhompiyutha eziku-inthanethi ngamabhizimusi amakhulu kuya kusanda.

Sekukonke, ezinhlanganisweni labathuthukisi abafuna i-AI yezinga eliphambili kakhulu evulekileyo, esebenza ngobungcwethi njalo ezimeleyo ngo-2026, i-Zyphra ingenye yezinkampani eziqakatheke kakhulu njalo ezithembisa kakhulu ezisebenza kuzo zonke izingxenye zalumkhakha wobuchwepheshe. Kayisiyo enye nje indawo yokucwaninga ngamamodeli — yakha ingqalasizinda epheleleyo yesikhathi esizayo yobuhlakani obuvulekileyo obudlula obabantu.

Zuza lomvuzo we-QUA nge-Quasa!  

4.8/5 izinkanyezi (ivelele ekuletheni okutsha, ekusebenzeni ngobungcwethi, lasekugxileni ekuzimeleni; kulezinto ezincane eziphawulwayo ngokungakavuthwa kwamamodeli amatsha asesebangeni lokuqala).

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