Zyphra: Uphononongo lweemodeli ze-MoE ezisebenza ngokufanelekileyo, eziqeqeshwe kwi-AMD nezigqwesa oko kulindelekileyo kubungakanani bazo

#Quasa #QUA #zyphra

I-Zyphra (eboniswe ku-Quasa.io/projects/zyphra) yinkampani yobukrelekrele obungaphaya kobomntu obuvulelekileyo, esebenza kuwo onke amanqanaba obugcisa, eyakha iinkqubo ze-AI ezizimeleyo, eziselubala nezihambelana neenjongo zabantu ukusuka kwisiseko.

Endaweni yokuxhomekeka kwiimodeli ezivaliweyo nakwizixhobo zekhompyutha zomthengisi omnye, i-Zyphra iphuhlisa iimodeli ezisisiseko ezintsha nezivulelekileyo (usapho lwe-ZAYA) ezinokuxinaniswa kobukrelekrele okuzisa inkqubela enkulu, ize izisebenzise nge-Zyphra Cloud — inika imibutho ulawulo lokwenene kwiziseko zayo ze-AI.

Eyona nto iphambili kwi-Zyphra kukutyhala imida yokusebenza ngokufanelekileyo kobukrelekrele: idala iimodeli ezikwaziyo ukuqiqa kwinqanaba lezona ziphambili ngoxa zisebenzisa iiparamitha ezisebenzayo ezimbalwa kakhulu kwaye zisebenza ngokufanelekileyo kwezinye iintlobo zeetshiphu.

Ezona zinto zibalulekileyo ziquka:
• ZAYA1-8B — Imodeli ye-Mixture-of-Experts eneep حaramitha ezisebenzayo eziyi-760M kuphela (eziyi-8.4B zizonke), egqwesa kakhulu oko kulindelekileyo kubungakanani bayo kwiimvavanyo zokulinganisa ukuqiqa okuntsonkothileyo, imathematika nokubhala ikhowudi — idla ngokulingana okanye yoyise iimodeli ezinkulu kakhulu ezinobunzima obufumaneka esidlangalaleni.
• Ulwakhiwo olutsha lwe-MoE++ — I-Compressed Convolutional Attention, ukuhanjiswa kwedatha okukwinqanaba eliphezulu, nezinto ezintsha kuqeqesho lwangaphambili/nasemva koqeqesho ezandisa ukusebenza nge-FLOP nganye kangangoko kunokwenzeka.
• Uqeqesho olubeka i-AMD kuqala — Iimodeli zokuqala ezinkulu ze-MoE eziqeqeshwe kwangaphambili, zaqeqeshwa phakathi, zaza zalungiselelwa imisebenzi ethile ukusuka ekuqaleni ukuya ekugqibeleni kwii-GPU ze-AMD Instinct MI300X nakwizixhobo zothungelwano ze-Pensando — zibonisa ukuba iziseko ze-AI ezizimeleyo, ezisebenzisa abathengisi abaninzi, zikulungele ukusetyenziswa.
• Zyphra Cloud — Iqonga lokusetyenziswa kwimveliso eliquka onke amanqanaba obugcisa, elizisa izinto ezintsha ze-Zyphra Research kubaphuhlisi, kumashishini nakubaboneleli beenkonzo zelifu abakhulu kakhulu, ngokusetyenziswa nokwandiswa komthamo okungenazithintelo.
• Iimodeli ezisebenzisa iintlobo ezininzi zedatha neemodeli ezizayo — I-ZAYA-VL (umbono nolwimi), iinguqulelo zovavanyo ze-ZAYA1-74B, iimvavanyo zeemodeli zolwimi ezisebenzisa ukusasazeka, nokunye okukwisicwangciso sophuhliso.

Ifaneleke kakhulu kubaphandi be-AI, kumashishini afuna i-AI ezimeleyo, kubaphuhlisi abakha kwiimodeli ezivulelekileyo, nakuye nabani na odiniweyo kukubotshelelwa kumthengisi omnye okanye ziindleko ezinkulu zokusebenzisa iimodeli ukwenza uqikelelo. Ngo-2026, i-Zyphra ibalasele njengenye yezona nzame zinika umdla zokuphuhlisa ubukrelekrele obungaphaya kobomntu obusebenza ngokufanelekileyo, obuvulelekileyo nobungaxhomekekanga kuhlobo oluthile lwezixhobo zekhompyutha.

Uluntu luyayithanda:
“I-ZAYA1-8B iyamangalisa — ineep حaramitha ezisebenzayo ezingaphantsi kwe-1B kodwa igqwesa kwiimvavanyo zokulinganisa imathematika nokubhala ikhowudi ezisokolisayo kwiimodeli ezinobungakanani obuyi-10x obayo. Eli likamva lobukrelekrele obusebenza ngokufanelekileyo.”
“Ekugqibeleni, kukho enye indlela yokwenene kwihlabathi elixhomekeke kwi-NVIDIA kuphela. Iimodeli ze-Zyphra eziqeqeshwe kwi-AMD zibonisa ukuba i-AI ezimeleyo ayinakwenzeka nje kuphela — sele ilapha.”
“Zyphra Cloud + iimodeli ze-ZAYA = uluhlu olupheleleyo lwetekhnoloji yobukrelekrele obungaphaya kobomntu obuvulelekileyo ebesisoloko silulindile.”

Ibalasela ngakumbi ekuxinaniseni ubukrelekrele, kwimisebenzi efuna ukuqiqa kakhulu, ekusebenziseni iimodeli ukwenza uqikelelo ngeendleko ezifikelelekayo, nasekwakheni iinkqubo ze-AI ezizimeleyo ngokwenene ngaphandle kokunciphisa ukusebenza.

Iintsilelo: Iimodeli zisekumanqanaba okuqala ophuhliso (iinguqulelo ezintsha ezifana ne-ZAYA1 ziinguqulelo zovavanyo); iziphumo ezizezona zilungileyo zidla ngokufuna izibalo ezongezelelweyo ngexesha lovavanyo ukuze kufikelelwe kowona mgangatho uphezulu wokusebenza; ukusetyenziswa kweqonga lelifu elipheleleyo kumlinganiselo wamashishini kuya kusanda.

Lilonke, kwimibutho nabaphuhlisi abafuna i-AI ekwinqanaba eliphambili, evulelekileyo, esebenza ngokufanelekileyo nezimeleyo ngo-2026, i-Zyphra yenye yezona nkampani zibalulekileyo nezithembisayo ezisebenza kuwo onke amanqanaba obugcisa kweli candelo. Ayiyiyo nje enye ilabhoratri yeemodeli — yakha iziseko ezipheleleyo zobukrelekrele obungaphaya kobomntu obuvulelekileyo bexesha elizayo.

Fumana nomvuzo we-QUA nge-Quasa!  

Iinkwenkwezi eziyi-4.8/5 (ibalasele kwizinto ezintsha, ekusebenzeni ngokufanelekileyo nasekugxileni ekuzimeleni; kukho amanqaku amancinane malunga nokuvuthwa kweemodeli ezintsha ezisekumanqanaba okuqala).

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