Ngeli xesha lobukrelekrele bokwenziwa, ukuvela kweemodeli ezintsha zokuvelisa kuyaqhubeka ukumangalisa abasebenzisi, abaphuhlisi kunye namashishini ngokufanayo. Iingxilimbela ezimbini zesimo sangoku, Gemma 3 y DeepSeek, balwela ubunkokeli betekhnoloji kunye nokukhethwa kwabo bafuna izisombululo ezisebenzayo nezinamandla ze-AI, kusetyenziso lomntu siqu kunye nobungcali.
Zombini iimodeli zenze umdla omkhulu kuluntu, nganye ineendlela ezahlukeneyo zokuvuleleka, ukusebenza kakuhle, kunye namandla okusebenza. Ngelixa enye ivela koomatshini bokusungula bakaGoogle, enye kukuzibophelela kweshishini laseTshayina ekwenzeni idemokhrasi i-AI ngaphandle kokuncama ukusebenza. Apha ngezantsi, siphonononga iimpawu, iingenelo, kunye nemida yeGemma 3 kunye ne-DeepSeek nzulu ngohlalutyo olubanzi, olufikelelekayo olulungiselelwe abaphuhlisi kunye nabasebenzisi abanomdla.
Yintoni iGemma 3?
IGemma 3 imele isizukulwana samva nje seemodeli ze-AI ezivulekileyo eziphuhliswe nguGoogle. Ngokungafaniyo neyakhe ivaliwe Gemini usapho, apho i-Google kuphela ikwazi ukufikelela kwikhowudi yayo, i-Gemma iyingxenye yomzamo wokuvula ukuphuhliswa kwayo kuluntu. Ngoyilo lomthombo ovulekileyo, iGemma 3 ifuna ukutsala abaphuhlisi, abaphandi, kunye neenkampani ezinomdla wokuphumeza i-AI ngaphandle kokuxhomekeka kwiinkonzo zangaphandle okanye iindleko eziphezulu.
Yintoni eyenza iGemma 3 ibe yodwa kukuguquguquka kwayo kunye nokufikeleleka kwayo. Ifumaneka kwiinguqulelo ezininzi ezinokulinganiswa ngokuxhomekeke kwinani leeparamitha, ukusuka kwiimodeli ezineebhiliyoni zeebhiliyoni zeeparitha ukuya kwezona ziphambili kunye ne-1.000 yeebhiliyoni zeeparamitha. Obu bubanzi buyenza ibe luncedo kwiimeko ezahlukeneyo, ukusuka kwimisebenzi elula yeselula ukuya kwizicelo ezintsonkothileyo zeshishini okanye zemfundo.
Iimpawu zobugcisa eziqaqambileyo zeGemma 3
Elinye lawona mandla makhulu eGemma 3 kukuba inokuqhuba nakwisixhobo esineGPU enye.. Oku kubeka imodeli kaGoogle kwindawo enqabileyo xa kuthelekiswa nabo bakhuphisana nabo, abathi bafuna amaziko edatha omelele ngakumbi okanye iziseko zophuhliso.
Ixhasa iilwimi ezingaphezu kwama-140, okwenza enye yeemodeli ezilungiselelwe kakhulu kwiimeko zeelwimi ezininzi. Oku kuquka iilwimi ezisetyenziswa kakhulu nezingaxhaphakanga kangako, nto leyo eyenza ukufikelela kwayo kwihlabathi jikelele kuphawuleke ngokwenene.
Ukongeza kwisicatshulwa, iGemma 3 iyakwazi ukucubungula imifanekiso kunye neevidiyo ezimfutshane.. Lo msebenzi we-multimodal wandisa kakhulu usetyenziso, ukusuka ekuhlalutyeni umxholo wemultimedia ukuvelisa iimpendulo ezibonwayo.
Into ephawulekayo yifestile yemeko yayo ukuya kuthi ga kwi-128.000 yamathokheni, ekuvumela ukuba usebenze ngamaxwebhu amakhulu okanye idatha eninzi edibeneyo, ebalulekileyo kwimisebenzi efana nezishwankathelo ezide, uhlalutyo olunzulu, okanye ixesha elide, iingxoxo zamanzi.
Izinto ezintsha zobugcisa eziphucula iGemma 3
IGemma 3 ibandakanya uyilo olusekwe kwi-transformer kunye nedekhowuda esulungekileyo, ilungele ukwenziwa kokubhaliweyo. Ukhathalelo lwenziwe ngendlela exubileyo nge imigangatho emihlanu yasekhaya kunye nelinye ilizwe jikelele, ukuvumela ukusebenza kakuhle ukuba kugcinwe ngaphandle kokuncama ukuqonda ukuxhomekeka kwexesha elide.
Imodeli ibandakanya i-encoder ebonakalayo eguqula imifanekiso ibe ngamathokheni ahambelanayo, ukuvumela okubhaliweyo kunye nomfanekiso ukuba kuqhutywe kunye kakuhle. Olu tshintsho olubonakalayo longeza izakhono zayo ezininzi.
Kwakhona, Ubungakanani busetyenziselwa ukunciphisa ubungakanani bemodeli ngaphandle kokubeka esichengeni ukusebenza kakhulu., kunye nezinye iindlela zobuchule ezinje ngegrouped query attention (GQA), ulwazi lwe-distillation olusuka kwimifuziselo ekwinqanaba eliphezulu, kunye nekopus eqeqeshelwe iilwimi ezininzi.
Enye inqaku eliphambili yi usebenzisa uhlengahlengiso lwengxelo yomntu (RLHF/AR), ekuvumela ukuba ulungelelanise kakuhle indlela yokuziphatha yomzekelo kwimisebenzi enje Ukucwangcisa, imathematika, ukuqiqa okunengqiqo, kunye nokuvelisa iimpendulo ezithembekileyo.
Izinto ezilungileyo zeGemma 3 ngaphezu kweemodeli ezivaliweyo
IGemma 3 yenye indlela enamandla ngakumbi ukuba ufuna ukuzimela kunye nolawulo olupheleleyo.. Ikhowudi yayo yomthombo ovulekileyo kunye nobukhulu obufikelelekayo buvumela ukuba iqhube kwiindawo zendawo ngaphandle kwesidingo se-APIs yorhwebo, iyenza ibe yinto efanelekileyo kuphuhliso lwabucala, lwemfundo, okanye ngaphandle kwe-intanethi.
Ukuqhuba imodeli ekuhlaleni kuphucula ubumfihlo kwaye kunciphisa amaxesha okubambezeleka. Enkosi kuyilo lwayo oluphuculweyo, kunokwenzeka ukuba uyisebenzise kwilaptop ngaphandle kweGPU okanye uyisebenzise kwizisombululo ezizinzisiweyo ngeGoogle AI Edge.
Ukongeza kuloo nto, uGoogle uphucukile Ukhuseleko ngohlelo lweShieldGemma 2, ehluza ngokufanelekileyo imifanekiso ecacileyo okanye enobundlobongela. Oku kuluncedo ngakumbi kwizicelo ezijolise kubaphulaphuli abahlukeneyo okanye iindawo ezingqongileyo ezinemigangatho ephezulu yokukhuselwa komxholo.
Uthelekiso lwentsebenzo: Ngaba iGemma 3 iyayigqwesa i-DeepSeek?
Kwiimvavanyo ezahlukeneyo zebenchmark kunye namaqonga ovavanyo angathathi hlangothi afana I-LMSYS Chatbot Arena, Gemma 3 ifumene iziphumo ezikhuphisana kakhulu, kwaye nkqu Ineemodeli ezigqwesileyo ezifana ne-LLaMA-405B kunye ne-DeepSeek-V3 kumgangatho wokuphendula olinganiswe ngumntu..
Ngokukodwa Inguqulo ye-parameter ye-27.000 yebhiliyoni ye-Gemma 3 iphumelele ukulinganisa kwe-Elo ye-1338, xa kuthelekiswa neemodeli ezisebenzisa ngaphezulu kokuphindwe kabini okanye kathathu iiparamitha. Oku kuyamangalisa xa kucingwa ukuba imodeli inokusebenza kwindawo enezixhobo ezincinci.
Kwiimvavanyo zemfundo ezifana ne-MATH kunye ne-MMLU-Pro, ifumene i-89 kunye ne-67,5 amanqaku ngokulandelanayo, ivelele kwingqiqo ehleliweyo, iprogram kunye nemisebenzi yokuqonda okuphambili.
Ngokubonakalayo, Isebenze kakuhle kwiibenchmarks ezifana ne-TextVQA kunye ne-InfoVQA, nangona isawela emva kweemodeli ezivaliweyo ezifana ne-GPT-4V. Nangona kunjalo, iimpendulo zabo zazisoloko zibonakaliswa ngokuchaneka ngokwemeko kwaye zihambelana negalelo elibonakalayo elibonelelweyo.
Yintoni i-DeepSeek kwaye yintoni eyenza ibe yinto ekhethekileyo?
I-DeepSeek iye yafumana ukuduma njengenye yezona AIs ezinamandla zemvelaphi yaseTshayina, evelele ngakumbi ngo-2024 kunye no-2025.. Imodeli yayo eyaziwa kakhulu, DeepSeek R1, yaphuhliswa eHangzhou kwaye ibonakaliswe ngokunika umsebenzi obalaseleyo odityaniswe nokusebenza kakuhle kunye nedemokhrasi enkosi kubume bayo obuvulekileyo.
I-DeepSeek ibonelela ngeendlela ezimbini eziphambili: i-R1, ejoliswe ekuqiqeni, kunye V3, iguquguquka ngakumbi kwimisebenzi eqhelekileyo. Zombini ezi nguqulelo zingasetyenziswa kuphuhliso lwasimahla noluhlawulweyo, umahluko bubude bomxholo kunye namandla ekhompyuter.
DeepSeek izakhono eziphambili
I-DeepSeek inokufikelela kwi-Intanethi, ilayishe iifayile kwaye ihlalutye umxholo, kunye nokukwazi ukwenza ikhowudi, imathematika ephucukileyo kunye nokuqiqa okunengqiqo. Yenzelwe imisebenzi yobuchwephesha kunye nenzululwazi, iyenza ibe lukhetho olukhulu kwiiprofayili zezemfundo kunye nezobuchwephesha.
Enye yeenzuzo zayo eziphawulekayo kukuba inokukhutshelwa ukusebenza ngaphandle kweintanethi, eyinxalenye yokhuseleko kunye neqhinga labucala apho abaphuhlisi banolawulo olungaphezulu kwidatha.
Ukusebenza kwayo kwimisebenzi yokuqiqa ibalasele, nto leyo ekhuthaze iinkampani ezininzi ukuba ziyamkele njengenxalenye yeenkqubo zazo zangaphakathi okanye abancedisi abakhethekileyo.
Umahluko ophambili phakathi kweGemma 3 kunye ne-DeepSeek
Zombini iimodeli zinokufana okuthile, njengokufumaneka njengomthombo ovulekileyo kunye nokuvumela ukwenziwa kwendawo, kodwa kukho iiyantlukwano ezininzi eziphawula indlela yazo kunye nokusetyenziswa kwamatyala:
- IGemma 3 ikhaphukhaphu kwaye iguquguquka kwihardware yamandla aphantsi, ngelixa i-DeepSeek ifuna izixhobo ezininzi zekhompyutha.
- IGemma 3 ixhasa ngaphezu kweelwimi ezili-140. kwaye i-DeepSeek igxile kusetyenziso lobuchwephesha, nangona ikwayilwimi ezininzi.
- I-DeepSeek yenzelwe ukuqiqa okucwangcisiweyo, ingqiqo yemathematika kunye nemisebenzi enzima, nangona i-Gemma 3 idibene okanye yadlula kwiimvavanyo ezininzi.
- IGemma ibonelela ngenkxaso yemultimedia enobuchule bokubonwayo phambili ngakumbi, ukudibanisa imifanekiso kunye neevidiyo ezimfutshane kwiinkqubo zabo.
IGemma 3 kwizixhobo zokwenyani: indlela yokuyisebenzisa
IGemma 3 inokuvavanywa ngokulula kwisikhangeli kwiGoogle AI Studio., isixhobo se-intanethi esingafuni uqwalaselo olongezelelweyo. Ikwafumaneka kumaqonga afana noGoogle Colab, uBuso obuHugging, iKaggle, kunye neOllama.
Enye yeenzuzo ezinkulu kukuba i-Ollama ikuvumela ukuba uqhube imodeli ekuhlaleni ngaphandle kweGPU., eququzelela uphuhliso lweendawo ezingaxhunyiwe kwi-intanethi ngokupheleleyo. Esi sisisombululo esifunwa kakhulu kwabo basebenza kunxibelelwano oluphantsi okanye kwiindawo eziyimfihlo.
Ukusuka kwizixhobo eziphathwayo, IGemma 3 inokudityaniswa neGoogle AI Edge, ukuvula ithuba lokusebenzisa i-AI ngaphandle kwe-intanethi kwimisebenzi efana nokutolika kwemifanekiso, ukuveliswa komxholo okhawulezayo, kunye nohlalutyo lombhalo.
Uthelekiso olusebenzayo: Yeyiphi onokuyikhetha ngokuxhomekeke kwimeko?
Zombini iimodeli zigqame kwiinkalo ezahlukeneyo. Ukuba ufuna imodeli ye-AI efikelelekayo, esebenzayo, ehambelanayo neelwimi ezininzi kwaye ekwaziyo ukusebenza ngokuzimeleyo kwilifu, iGemma 3 ingaphezulu kokubheja okuqinileyo.
Ngakolunye uhlangothi, Ukuba ugxininiso lunobugcisa ngakumbi kwaye ufuna imodeli ephendula ngokuchanekileyo iingxaki zemathematika, inkqubo okanye inferences yesayensi enzima., DeepSeek R1 inokuba lolona khetho lulungileyo.
Ngokumalunga nokuguquguquka, iGemma 3 inenzuzo ngokukwazi ukusetyenziswa kwi-GPU okanye ngaphandle kwayo, okunciphisa imiqobo yokungena. I-DeepSeek, nangona inamandla, idinga iimfuno eziphezulu kancinci ukuze iqhube kakuhle.
Kwiimvavanyo ezingaboniyo kunye nebenchmarksIGemma 3 igqwesile kuzo zonke iimetriki ezininzi eziphambili, ebonisa ukukhula kwayo njengesisombululo esibanzi se-AI ngo-2025.
Oko yasungulwa, IGemma 3 iguqule ukufikelela kubukrelekrele bokwenziwa bokusebenza okuphezulu. ngokugxila ekusebenzeni kakuhle, ukhuseleko kunye nokuvuleleka. Ngelixa i-DeepSeek ihlala iyibhenchmark ngokwemigaqo yamandla okuqiqa kunye nengqiqo yobugcisa, isindululo sikaGoogle sibonelela ngesisombululo esilungeleleneyo ngakumbi phakathi kwamandla, ukufikeleleka, kunye nezicelo zehlabathi lokwenyani.
Zombini ezinye iindlela zibonelela ngeendlela eziqinileyo, kodwa ukuba ujonge izinto ezininzi, zikhaphukhaphu, kunye ne-AI etyebileyo, iGemma 3 ngokuqinisekileyo lolona khetho lunyanzelisayo kwinkqubo yanamhlanje evulekileyo ye-AI. Yabelana ngolwazi ukuze abantu abaninzi bazi malunga nesihloko..