Gemini 3.7 Flash Uyasebenza

I-Gemini 3.7 Flash iyimodeli entsha eyinhloko ye-AI eyethulwe yi-Google DeepMind. Le modeli yakhelwe ngokukhethekile izinqubo ze-Coding ne-Agent, futhi ithuthuke kakhulu kumabhentshimakhi obunjiniyela besofthiwe, ukuthuthukiswa kwamakhasi ewebhu nokuzenzakalela kwamabhizinisi, kanti amanye amamethrikhi adlula i-Claude Sonnet 5 ne-GPT-5.6 Terra. Intengo yokuqala yale modeli inguhhafu kuphela wentengo yokuqala ye-3.6 Flash, futhi ibekwe njengemodeli enenani elihle kakhulu yezinqubo ze-Agent ezisetshenziswa njalo nezinqubo zamabhizinisi.

Ubunjiniyela besofthiwe: Isekela ukukhiqizwa kwekhodi yekhwalithi ephezulu elungele ukukhiqizwa, futhi ithuthuke kakhulu kumabhentshimakhi e-FrontierCode ne-DeepSWE, ngokuthuthuka okuphawulekayo ekunembeni kokuqala nasemandleni okulungisa amaphutha.
Ukuthuthukiswa kweWebhu: Isekela ukufaka izithombe-skrini, izithombe noma i-Design System ephelele ukuze ikhiqize amakhasi ewebhu asebenzisanayo anokwethembeka okuphezulu ngokuchofoza okukodwa, futhi ihola izimbangi ngesilinganiso se-Code Arena Elo.
Ukuzenzakalela kwe-Agent: Isebenzisa izinqubo zamabhizinisi eziyinkimbinkimbi ezinezinyathelo eziningi, ihlele ngokuzenzakalela, ibize amathuluzi futhi izivumelanise nezimo lapho ihlangabezana nezithiyo, ngokuthuthuka okukhulu ku-AutomationBench.
Umsebenzi odinga ulwazi oluningi: Iphatha ukucabanga okuyinkimbinkimbi kwemibhalo emikhakheni yezezimali, ezomthetho nesayensi yezinto eziphilayo, ngokuthuthuka okuphawulekayo ekuqondeni imibhalo emide efana ne-GDP.pdf.
Ukudala okuhlanganisa izindlela eziningi: Ngokuhambisana ne-Nano Banana, iguqula iziqondiso zombhalo zibe imidlalo ye-3D edlalekayo ngesikhathi sangempela, noma iguqule i-PDF emile ibe ikhasi lewebhu eliguqukayo elinamashadi asebenzisanayo.
Ukuhlanganiswa kwe-Gemini Spark: Njengemodeli esebenza ngemuva ye-Agent yomuntu siqu esebenza 24/7, iqhuba imisebenzi yokuzenzakalela ngaphakathi kwe-Google Workspace, njengokuhlela amafayela, ukubhala okusalungiswa kwama-imeyili nokubuyekeza izimo.

Ukuphindaphindwa kwama-algorithm nendlela yokuthuthuka okusheshayo: Imodeli ithuthukiselwe ngemuva kokuqeqeshwa futhi yaqiniswa ngokugxila emandleni ezimo ze-Coding ne-Agent, ngaphandle kokuthembela kuphela ekwandiseni usayizi wokuqeqeshwa kwangaphambilini. Ngaleyo ndlela igcina inzuzo yesivinini sochungechunge lwe-Flash, kuyilapho ithuthukisa kakhulu ikhwalithi yobunjiniyela besofthiwe nomsebenzi odinga ulwazi.
Ukuhlela okunezinyathelo eziningi nokusebenzisa ukucabanga okuqinile: Imodeli inendlela esebenzayo yokwaba izinsiza zokucabanga. Lapho ibhekene nomsebenzi oyinkimbinkimbi, inikeza izinsiza eziningi zokubala ekuhleleni okunezinyathelo eziningi nasekubizeni amathuluzi, futhi izivumelanise namasu ayo lapho ihlangabezana nezithiyo, icacise nenhloso. Lokhu “kusebenza okuqinile” kunciphisa kakhulu isidingo sokuqondiswa umuntu kanye nenani lokuzama futhi, futhi kuyisisekelo sobuchwepheshe bokuthuthuka kwamandla ayo e-Agent nokuzenzakalela kwamabhizinisi.
Ukuqonda okuhlanganisiwe kwasekuqaleni kwezindlela eziningi: I-3.7 Flash inamandla okuqonda ahlanganisiwe kuwo wonke umbhalo, umsindo, izithombe, ikhodi nevidiyo. Lokhu kuyenza ikwazi ukuhlanganisa nokucubungula okokufaka okubukwayo, njengezithombe-skrini nemidwebo yokuklama, kanye nemiyalelo yolimi ezimweni ezifana nokukhiqizwa kwemidlalo ye-3D, umjikelezo wokuqeqeshwa kwamarobhothi nokuguqulwa okuhlakaniphile kwemibhalo. Ngaleyo ndlela ikhiqiza yonke into kusukela ku-PDF emile kuya ekhasini lewebhu elisebenzisanayo, kanye nasencazelweni yombhalo kuya empahleni ye-3D esebenzayo.
Izinqubo ze-Agent nokuhlelwa kwama-Sub-Agent: Yakhelwe ngaphansi ngokukhethekile izimo ze-Agent ezisetshenziswa njalo. Ngokuthuthukisa ukunemba kokulandela imiyalelo kanye nekhono lokuzivumelanisa nezithiyo, isekela imodeli eyinhloko ekuphatheni ama-sub-agent amaningi ukuze aqedele imisebenzi ngokubambisana. Leli khono lokuhlela liyenza ikwazi ukucubungula ngokufana ukukhiqizwa kwesakhiwo, ukubhala umqondo wokusebenzisana nokunikezwa kwezingxenye ezibukwayo ezimweni ezifana nokuthuthukiswa kweWebhu, ibe ingqalasizinda engemuva yokuzenzakalela kwamabhizinisi esekela ukusebenza isikhathi eside nokusetshenziswa ngezinga elikhulu.

Ukuxhuma konjiniyela: Onjiniyela bangabiza imodeli ngokuqondile nge-Gemini API noma ku-Google AI Studio.
Ukuhlola okusheshayo: Ku-Google AI Studio ungakha iphrojekthi mahhala bese ukhetha imodeli ye-gemini-3.7-flash ukuze uhlole ngokushesha amakhono ombhalo, ekhodi kanye nezindlela eziningi.
Ukuhlanganiswa kweselula: I-Android Studio isivele ihlanganise le modeli, okuvumela onjiniyela ukuthi bathole usizo lokubhala nokukhiqiza ikhodi enqubweni yokuthuthukisa izinhlelo zokusebenza zeselula.
Ukuhlela i-Agent: Nge-Google Antigravity, indawo yokuthuthukisa egxile ku-Agent, ungahlela ama-sub-agent ukuze wakhe izinqubo eziyinkimbinkimbi zokuzenzakalela.
Ukusetshenziswa kwamabhizinisi: Abasebenzisi bamabhizinisi bangasebenzisa i-Gemini Enterprise Agent Platform ukwenza ngokwezifiso izinqubo zebhizinisi nokuhlanganiswa kwamathuluzi, ukuze basebenzise imodeli njalo ngezinga elikhulu.
Ukusetshenziswa kwabasebenzi: Abasebenzi bamabhizinisi bangasebenzisa ngokuqondile i-Gemini Enterprise App njengendawo yokungena eyodwa, basebenzise amandla akhelwe ngaphakathi e-3.7 Flash ukuphatha imibhalo nemisebenzi yokuzenzakalela.

Isivinini esikhulu sokuthuthuka: Uchungechunge lwe-Flash luthuthuke lusuka ku-3.5 lwaya ku-3.7 ezinyangeni ezi-3 kuphela, kanti ukusuka ku-3.6 kuya ku-3.7 kuthathe amasonto ama-3 kuphela. Ijubane i-Google ephendula ngalo impendulo yonjiniyela nokuncintisana kwemakethe alifaniswa namanye amamodeli amakhulu.
Ukuthuthuka okukhulu kwe-Coding: I-FrontierCode 1.1 Main inyuke isuka ku-34.4% yaya ku-43.6%, kanti i-DeepSWE v1.1 yagxuma isuka ku-49% yaya ku-65.3%. Ukunemba kwekhodi ekuzameni kokuqala kanye nekhono lobunjiniyela besofthiwe besikhathi eside kuthuthuke kakhulu, futhi ikhwalithi yokukhiqiza ikhodi manje isezingeni lemodeli eyinhloko elungele ukukhiqizwa.
Ukuhola okuphelele ku-Agent nokuzenzakalela kwamabhizinisi: I-AutomationBench ithole u-30.4%, cishe okuphindwe kabili kwesizukulwane sayo sangaphambilini, futhi idlula kakhulu i-Claude Sonnet 5 (10.7%) ne-GPT-5.6 Terra (23.6%). Lokhu kuyenza ibe yimodeli enamandla kakhulu okwamanje ezimweni zokuzenzakalela kwezinqubo zamabhizinisi.
Ukudlula izimbangi ezisezingeni eliphezulu kwamanye amabhentshimakhi: Kumabhentshimakhi amabili abalulekile e-Coding, i-FrontierCode 1.1 (43.6%) ne-WebDev Code Arena (1588 Elo), i-3.7 Flash isivele idlule i-Claude Sonnet 5 ne-GPT-5.6 Terra, yaphula umbono wokuthi “i-Flash yenzelwe usizo olulula kuphela”.
Inani elihle kakhulu: Intengo yokuphromotha ingu-$0.75 wokufaka no-$3.75 wokukhipha kuma-Tokens ayisigidi, okuyingxenye kuphela yentengo yokuqala ye-3.6 Flash, futhi iyingxenye eyodwa kwezintathu kuya kweyesine yentengo ye-Claude ($2/$10) ne-GPT ($2/$12). Ezinhlelweni ze-Agent ezisetshenziswa njalo nasezinqubweni zamabhizinisi, le nzuzo yezindleko ikhula kakhulu ngokwanda kokusetshenziswa.
Ukudala okuhlanganisa izindlela eziningi kusukela ekuqaleni kuya ekugcineni: Isekela ukukhiqiza imidlalo ye-3D edlalekayo ngombhalo ngokuqondile, ukuguqula izithombe-skrini noma i-Design System zibe amakhasi ewebhu asebenzisanayo ngokuchofoza okukodwa, kanye nokuguqula ama-PDF amile abe izindaba zedatha eziguqukayo ezinamashadi esikhathi sangempela. Inqubo yokudala iphelele futhi ayidingi ukuhlanganiswa kwamamodeli amaningi.

Iwebhusayithi yephrojekthi: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/

Umsizi ohlakaniphile we-Coding: Ukhiqiza ngokuzenzakalela ikhodi elungele ukukhiqizwa, ulungise ama-Bug ayinkimbinkimbi futhi uqedele imisebenzi yobunjiniyela besofthiwe yesikhathi eside, ngokuthuthuka okukhulu ekunembeni kokuqala uma kuqhathaniswa nesizukulwane sangaphambilini.
Ukuthuthukiswa kweWebhu okusebenzisanayo: Ukhiqiza ikhasi lewebhu elisebenzisanayo nelinokwethembeka okuphezulu ngesithombe-skrini noma i-Design System ngokuchofoza okukodwa, futhi isekela ukubambisana kwama-agent amaningi nemiphumela yokunyakaza ye-parallax.
Ukuzenzakalela kwezinqubo zamabhizinisi: Iqhuba i-RPA nezinhlelo zebhizinisi, iqedele ngokuzenzakalela imisebenzi ephindaphindwayo efana nokufaka idatha ezinkundleni ezahlukene, ukudlulisa izicelo zokugunyazwa nokukhiqiza imibiko.
Ukukhiqizwa kwemidlalo ye-3D nokuqukethwe: Ngokuhambisana ne-Nano Banana, iguqula izincazelo zombhalo zibe imidlalo ye-3D edlalekayo ngesikhathi sangempela, ikhiqize ngamandla abalingiswa, izinto nemidwebo yezindawo.
Ukuguqulwa okuhlakaniphile kwemibhalo: Iguqula ngokuzenzakalela imibiko yonyaka ye-PDF emile nemibiko yocwaningo ibe amakhasi ewebhu aguqukayo noma izindaba zedatha ezinamashadi esikhathi sangempela, izifinyezo zedatha nemisebenzi yokusebenzisana.