I-Standard Intelligence ibhengeze ukupapashwa kwe-hertz-dev, imodeli yokuqala ye-AI evulekileyo ye-full-dupex intetho enokuthi isetyenziswe njengesiseko sokwakha unxibelelwano lwezwi lwexesha langempela okanye iinkqubo zokuvelisa ulwimi oluthethwayo. Imodeli inokuvelisa intetho efana ngokusondeleyo nedatha yelizwi apho iqeqeshelwa khona, ibonelela ngonxibelelwano lwendlela yomntu ngaphandle kokubambezeleka kwencoko yomnxeba. Uphuhliso lweprojekthi lusasazwa phantsi kwelayisensi ye-Apache 2.0.
Kwinkqubo ene-NVIDIA GeForce RTX 4090 GPU, i-avareji ye-pre-generation latency yi-120 ms (ngokwethiyori ukuya kwi-65 ms), ephantse ibe kabini ngokukhawuleza kuneemodeli ezikhoyo esidlangalaleni. Uguqulelo olupapashiweyo lwakhiwe ngokusebenzisa i-architecture yeTransformer, ihlanganisa i-8.5 yeebhiliyoni zeeramitha kwaye iqeqeshelwa ukusebenzisa i-500 yeebhiliyoni zamathokheni. Ubungakanani bomxholo othathelwe ingqalelo ngumzekelo (inani lamathokheni ukuba imodeli inokuyiqhuba kwaye ikhumbule xa ivelisa intetho) ngamathokheni angama-2048 okanye malunga nemizuzu emi-4 yokuthetha.
umthombo: opennet.ru
