Gemini 4 Carbon Uyasebenza

I-Gemini 4 Carbon iyimodeli enkulu yakamuva esezingeni eliphambili eyethulwe yi-Google DeepMind, futhi iyimodeli yesizukulwane esilandelayo esezingeni eliphambili ngemva kwe-Gemini 4 Argon. I-Gemini 4 Carbon ithathe indawo yokuqala ku-Artificial Analysis Intelligence Index 5.0 ngamaphuzu angu-59, yadlula i-Claude Opus 5.5 ne-GPT-6 Astra, futhi yafinyelela ezingeni elisha le-SOTA ekutholeni ubungozi, ekuqinisekiseni ukusetshenziswa kwabo, ekulungiseni iziqephu nasezabelweni zokuzihlela ezizimele zesikhathi eside kuwo wonke umjikelezo wokuvikela ukuphepha kwenethiwekhi.

Ukuhlela okuzimele kwesikhathi eside: Ingasebenza izinsuku eziningi ilandelana, ithole ngokuzimela imbangela eyisisekelo yama-Bug afihlekile kumakhosombe amakhulu ekhodi bese ibhala iziqephu zokuwalungisa.
Ukutholwa kobungozi: Ku-SEC-bench Pro ithole i-SOTA entsha ngo-88.2%, futhi ingathola ngokuzimela ubungozi obuyinkimbinkimbi nobukhulu ku-software yangempela efana ne-V8 engine ne-Linux kernel.
Ukuqinisekiswa kokusetshenziswa kobungozi: Ithuthukise irekhodi le-ExploitGym ngo-54.8%, iqinisekisa njengomhlaseli ukuthi ubungozi bungasetshenziswa ngempela, ukuze isize abavikeli bahlele kuqala ukulungiswa.
Ukufaka iziqephu okuzenzakalelayo ezingeni lembangela: I-CWE-bench ifinyelele ku-71.0%, futhi ingaguqula ngokuqondile ukutholakala kobungozi okuqinisekisiwe kube ikhodi yokulungisa esebenzayo nengaphazamisi imisebenzi ekhona.
Ukuthuthukiswa kwe-full-stack nokwakha i-frontend: I-WebDev Arena Elo inyuke ngamaphuzu angu-89 yafinyelela ku-1769, ikhwalithi yezakhi nemiphumela yokugcina yathuthuka kakhulu, kanti izindleko ezimaphakathi zomsebenzi zehla.
Ukuqulwa kwezinqumo ngezinyathelo eziningi ngaphansi komongo omude kakhulu: Ekuhlolweni kwe-GraphWalks okunomongo ongu-256k–1M ithole u-96.2% futhi yadlula kakhulu izimbangi, kanti ekulingiseni kwebhizinisi kwesikhathi eside i-Vending-Bench 2, ibhalansi yokugcina ye-akhawunti yacishe yaba izikhathi ezingu-1.8 kune-Opus 5.5.

Isakhiwo se-ejenti ezimele yesikhathi eside: I-Carbon yakhelwe ngokukhethekile “imisebenzi yobunjiniyela eqhubeka izinsuku eziningi enesikhathi eside sokuhlela”. Umnyombo wayo yindlela yokuhlola ephikelelayo lapho ibhekene namakhosombe amakhulu ekhodi nezinhlelo ezisezingeni eliphansi; ngisho noma ihlangana nemikhawulo kaningi, ingabuyela emuva, ishintshe indlela, iqhubeke nokuzama, futhi iqede ngokuzimela umjikelezo wezinyathelo eziningi kusukela ekutholeni imbangela eyisisekelo kuya ekulethweni kwesiqephu.
Ikhono lomjikelezo ogcwele wokuvikela ukuphepha kwenethiwekhi: Amakhono ayo okuphepha akhiwe ngezigaba ezintathu “thola ubungozi—buqinisekise—bulungise”, futhi zonke zaqeqeshwa zahlolwa zisesimweni esigcwele esihlukanisiwe emikhawulweni yokuvikela yomkhiqizo. Uhlangothi lokuthola lufinyelele ku-88.2% ku-SEC-bench Pro, uhlangothi lokuqinisekisa lwaphula irekhodi le-ExploitGym ngo-54.8%, kanti uhlangothi lokulungisa luhola i-CWE-bench ngo-71.0%; lezi zintathu zakha uchungechunge oluphelele kusukela ekutholeni kuya ekuvalweni komjikelezo. Idatha esemthethweni ikhombisa ukuthi izinga lokuphumelela lokukhiqiza ikhodi yokuphepha ngaphakathi lenyuke ngaphezu kokuphindwe ka-2 uma liqhathaniswa ne-Argon. Ekuqhathaniseni okwenziwa yithimba le-Chrome emsebenzini wangempela, ikhwalithi yeziqephu yayingaphezu kokuphindwe ka-10 kunemodeli esezingeni eliphambili evela kwenye ilabhorethri ehamba phambili.
Ukuqeqeshwa kokuqondanisa nokuncipha kokukopela: Ukuze kubhekwane nenkinga yamamodeli athola amaphuzu aphezulu “aguqule imibhalo yokuhlola ukuze azuze ukuphasa okungamanga”, i-Carbon icindezela ngokuqondile ukuthambekela kokusebenzisa izinqamuleli nezikhala ngesikhathi sokuqeqeshwa. Ngenxa yalokho, izinga lokukopela okuphumelelayo ku-Terminal-Bench 4.0 lehle laya ku-0.0%, izinga lokuzama ukukopela lehla laya ku-0.4%, kanti izexwayiso zokungaqondani kobukhulu obuphezulu ezindaweni zangaphakathi zokubhala ikhodi yesikhathi sangempela zehla cishe ngo-80% ngesikhathi ukunemba kuthuthuka.

Okwamanje i-Carbon ivulekele kuphela ingxenye yabasebenzi bokuvikela ukuphepha kwenethiwekhi abadlule ku-Fairwind Program. Abathuthukisi abajwayelekile, amabhizinisi nabathengi kufanele balinde isaziso esilandelayo esivela kwa-Google.

Ubuhlakani obuphelele buhamba phambili: Ithole indawo yokuqala emhlabeni ku-Artificial Analysis Intelligence Index 5.0 ngamaphuzu angu-59, yadlula ngqo i-Claude Opus 5.5 ne-GPT-6 Astra.
Ukuhlela okuzimele kwesikhathi eside: Ingasebenza izinsuku eziningi, idlule emikhawulweni, futhi ithole ngokuzimela imbangela yama-Bug afihlekile kumakhosombe amakhulu ekhodi bese ilungisa ngesiqephu; i-Terminal-Bench 4.0 ifinyelele ku-65.2% futhi yahola izimbangi.
I-SOTA kuwo wonke umjikelezo wokuphepha kwenethiwekhi: Ihlanganisa yonke inqubo “yokuthola ubungozi (SEC-bench Pro 88.2%)—ukuqinisekisa ubungozi (ExploitGym 54.8%)—ukufaka isiqephu (CWE-bench 71.0%)”, futhi yomithathu iyirekhodi lakamuva.
Ukunqoba komongo omude kakhulu: Ekuhlolweni kwe-GraphWalks okunomongo ongu-256k–1M ithole u-96.2%, yadlula kakhulu i-Opus 5.5 (66.8%) ne-Astra (71.8%).
Ukukopela okucishe kube yiziro kanye nokungazenzi (ezingeni lokuqondanisa): Izinga lokukopela okuphumelelayo ku-Terminal-Bench lehle laya ku-0.0%, kwaba nokuvikeleka okungu-100% ekujovweni kwemiyalelo engaqondile, kanti izitatimende ezingenabufakazi zehla cishe ngokuphindwe ka-2 uma kuqhathaniswa ne-Argon.

Iwebhusayithi esemthethweni yephrojekthi: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/

Ukuvikela ukuphepha kwenethiwekhi: Isebenzisa i-SEC-bench Pro engu-88.2%, i-ExploitGym engu-54.8% ne-CWE-bench engu-71.0% ukuze izenzele wonke umjikelezo “wokuthola ubungozi—ukuqinisekisa ubungozi—ukufaka isiqephu”.
Ukuhlela okuzimele kwesikhathi eside: Ingasebenza izinsuku eziningi idlule emikhawulweni, ithole ngokuzimela imbangela yama-Bug afihlekile kumakhosombe amakhulu ekhodi bese iwalungisa.
Ukuhamba komsebenzi kwama-ejenti asezingeni lebhizinisi: Ngenxa ye-Vals Index engu-70.4% kanye nesaphulelo esingu-95% kokokufaka okugcinwe kunqolobane, isekela imisebenzi yama-Agent ejwayelekile ngezindleko eziphansi.
Ukuhlaziywa komongo omude kakhulu: Iwindi lomongo elingu-1M nokusebenza kwe-GraphWalks okungu-96.2% kuyenza ifaneleke ukuqonda amakhosombe ekhodi nokwenza ukucabanga kwezinyathelo eziningi kumadokhumenti amade.
Ukuthuthukiswa kwe-Web ye-full-stack: Ngekhwalithi yemiphumela engu-1769 ku-WebDev Arena nezindleko eziphansi zomsebenzi, ikhiqiza ngokushesha izakhi ze-frontend nezinhlelo zokusebenza eziphelele.