Gemini 4 Argon Uyasebenza

I-Gemini 4 Argon iyimodeli yolimi olusezingeni eliphambili yesizukulwane esisha eyethulwe yi-Google DeepMind, egxile ekuhambeni komsebenzi okuyinkimbinkimbi nokuthatha isikhathi eside, okuhlanganisa umsebenzi wolwazi wamabhizinisi kwezobunjiniyela besoftware, ezezimali nezomthetho, kanye nokuvikela ukuphepha kwenethiwekhi. I-Gemini 4 Argon inyusa kakhulu umkhawulo wokukhishwa okukodwa, isuka kuma-Token angu-64K iye kuma-Token ayizigidi ezingu-1, futhi isekela ukuqedwa kwemisebenzi yesikhathi eside njengokuthuthwa kwamakhodi amakhulu ngesikhathi esisodwa. I-Gemini 4 Argon ithole u-77.9% ekuhlolweni kobunjiniyela besoftware kwe-DeepSWE v1.1, futhi ikhishwa ngezigaba ngohlelo lwe-Fairwind, iqale kubavikeli abathembekile bezokuphepha kwenethiwekhi.

Ukukhishwa okude kakhulu: Umkhawulo wokukhishwa okukodwa ukhushulwe kuma-Token angu-64K waya kuma-Token ayizigidi ezingu-1, okwenza ikwazi ukukhiqiza amakhulu ezinkulungwane zama-Token ekhodi ende noma imibiko ejulile ngesikhathi esisodwa ngaphandle kokuphazamiseka.
Ubunjiniyela besoftware: Ithole u-77.9% ekuhlolweni kwangempela kobunjiniyela besikhathi eside kwe-DeepSWE v1.1, futhi isekela ukulungisa amaphutha ekhodi, ukuklama ama-algorithm nokuthutha amakhodi amakhulu.
Umsebenzi wolwazi wamabhizinisi: Ibonisa ukusebenza okuhamba phambili emikhakheni yezezimali, ezomthetho nezentela, futhi isendaweni yokuqala kuzo zonke izigaba ze-Vals Index enesisindo esisekelwe ku-GDP yaseMelika.
Ukuqonda imidiya eminingi: Iyakwazi ukuhlaziya amashadi ochwepheshe, iqonde amavidiyo amade futhi ithathe izinyathelo isebenzisa imibhalo eminingi.
Ukuvikela inethiwekhi: Iyakwazi ukuthola, ukuqinisekisa nokulungisa ngokuzenzakalela ubuthakathaka obubalulekile. Ithole u-68% ku-CWE-bench v1 yabelana ngendawo yokuqala, futhi iyakwazi ukuthola izindawo zokuhlasela ekuhlolweni kokungena kwe-black-box ngaphandle kwekhodi yomthombo.
Ukwenza ngokuzenzakalela izinqubo zebhizinisi: Ibe sendaweni yokuqala ekuhlolweni kokwenza ibhizinisi ekupheleni kuya ekupheleni kwe-Zapier AutomationBench.
Ukuvikelwa kokuphepha: Inezingqimba ezine zokuvikela ezihlanganisa ukuvimbela ukusetshenziswa kabi, ukumelana nokufakwa kwemiyalelo engaqondile, ukuqapha ukungahambisani kochungechunge lokucabanga kanye nokuqiniswa kwe-sandbox.

Isakhiwo sokucabanga sesikhathi eside: Ngokusemthethweni, i-Argon inyusa umkhawulo wayo wokukhishwa usuka kuma-Token angu-64K uye kuma-Token ayizigidi ezingu-1. Uma imodeli ikwazi ukukhiqiza amashumi ezinkulungwane zama-Token emzileni womsebenzi owodwa, ingenza ukucabanga okujulile ngaphandle kokuphazamiseka, iqede imisebenzi eminingi eyinkimbinkimbi efana nokuthutha amakhodi amakhulu nocwaningo olunzulu ngesikhathi esisodwa.
Amakhono obunjiniyela asebenza njengama-agent: Imodeli isebenza ngokubambisana kwama-agent amaningi, ihlaziye idatha ye-telemetry futhi izenzele ngokwayo izinto ezifana nokuthuthukisa ukusetshenziswa kwememori nokubhala kabusha ikhodi ngemijikelezo eminingi yokuhlola ukusebenza. Konke ukuthuthwa kwamakhodi amakhulu kudinga ukuhlolwa okuzenzakalelayo, ukuhlolwa kokulingisa nokubuyekezwa kwabantu ngaphambi kokuthi kukhishwe.
Ukuvimbela ukusetshenziswa kabi nokufakwa kwemiyalelo: Imodeli yenqaba izicelo zokuhlasela inethiwekhi nokusetshenziswa kabi kwe-CBRN ngokwe-Frontier Safety Framework, kodwa igcina ucwaningo olusemthethweni olusetshenziswa ngezinhloso ezimbili. Iqinisa ukuqina kwayo ngokuqapha ukusebenza kwangaphakathi nokuhlolwa kwamaqembu abomvu, futhi imelane nokufakwa kwemiyalelo engaqondile ngamaqembu abomvu azenzakalelayo nokuqeqeshwa okuphikisanayo.
Ukuqapha ukungahambisani nokuqinisa uhlelo: Uchungechunge lokucabanga nezenzo ze-Argon kuqashwa kuyo yonke inqubo, futhi iyamiswa uma kutholakala ukuchezuka enjongweni yomsebenzisi. Inqubo yokuqeqesha nayo inesistimu yokuqapha efanayo ethumela izexwayiso eqenjini lokuphendula izehlakalo, kuyilapho imiphumela yokuqapha ingabuyiselwa ekuqeqesheni ngenhloso yokuvimbela ukugwema ukuqapha. Ngaphambi kokuqeqeshwa nokuhlolwa okunobungozi obukhulu, indawo ye-sandbox iyahlukaniswa futhi ivalwe ngokuphelele.

Qinisekisa ukufaneleka: Njengamanje i-Gemini 4 Argon ayikavulelwa wonke umuntu. Okokuqala kufanele uqinisekise ukuthi ungowaliphi iqembu labasebenzisi bokuqala: abavikeli abathembekile bezokuphepha kwenethiwekhi, amakhasimende e-API akhokhelayo noma ababhalisele i-Google AI Ultra.
Isiteshi sokufaka isicelo sabavikeli: Ochwepheshe bezokuphepha kwenethiwekhi bangafaka isicelo sokufaneleka kokuyizama kwasekuqaleni kuwebhusayithi esemthethweni ye-Google Fairwind Program https://deepmind.google/fairwind-program/. Ngemva kokwamukelwa, bazothola uhlobo oluphelele lokuvikela olungenazo izithiyo zenethiwekhi.
Ukubhalisa kwabasebenzisi abajwayelekile: Abathuthukisi nabasebenzisi bamabhizinisi kufanele balinde ukuvulwa kancane kancane okusemthethweni. Abasebenzisi bokuqala bangaxhuma nge-API ekhokhelwayo, kanti ababhalisele i-AI Ultra bazothola imvume yokusebenzisa kuqala.
Ukuxhuma nge-API: Ngemva kokuthola ukufaneleka, dala iphrojekthi ku-Google AI Developer Platform bese ubiza i-API ye-Gemini 4 Argon.

Ukukhishwa okude kakhulu: Umkhawulo wokukhishwa okukodwa ufinyelela kuma-Token ayizigidi ezingu-1, cishe uphindwe ka-8 kunamamodeli amaningi aphambili, acishe abe ngu-128K, futhi ungakwazi ukuqeda imisebenzi yesikhathi eside ngesikhathi esisodwa ngaphandle kokuphazamiseka.
Ubuholi kwezobunjiniyela besoftware: Ithole u-77.9% ekuhlolweni kwangempela kobunjiniyela besikhathi eside kwe-DeepSWE v1.1, yaphula irekhodi, futhi isekela ukuthuthwa kwamakhulu ezinkulungwane zemigqa yekhodi ye-C/C++ iye ku-Rust ngaphakathi kwe-Google.
Ukuhola okuphelele emsebenzini wolwazi wamabhizinisi: Isendaweni yokuqala kuzo zonke izigaba zezezimali, ezokuhlela, ezomthetho nezentela ku-Vals Index enesisindo esisekelwe ku-GDP yaseMelika.
Ukuqonda imidiya eminingi okuvelele: Ithole u-91.7% esilinganisweni sokuqonda amavidiyo amade se-LVBench, futhi iyakwazi ukuhlaziya amashadi ochwepheshe nokuthatha izinyathelo isebenzisa imibhalo eminingi.
Indawo yokuqala ekuzisebenziseni kwebhizinisi: Ibe sendaweni yokuqala ku-Zapier AutomationBench ekulinganisweni kokwenza ibhizinisi ekupheleni kuya ekupheleni, ngo-51.3%.
Ubuholi ekuvikeleni inethiwekhi: Yabelana ngendawo yokuqala ku-CWE-bench v1 ngo-68%, futhi yake yathola ubuthakathaka obuyingozi obuthinta isoftware yezokwelapha ezibhedlela emhlabeni wonke.

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

Ukuthutha amakhodi: Ithutha ngokuzenzakalela amakhodi angu-800,000 nangaphezulu e-C/C++ emitapo yolwazi esemqoka efana ne-Google re2 ne-libgav1, kanye ne-Fuchsia Zircon kernel, iwaguqulele ku-Rust ephephile kwimemori futhi iwakhuphe ngemva kokuhlolwa.
Ukubhala kabusha ikhodi ebalulekile ekusebenzeni: Ku-libgav1, i-decoder yevidiyo yomthombo ovulekile, ibhala kabusha ngokuzenzakalela imigqa engu-32,000 yekhodi ye-SIMD ibe yi-Rust ephephile engavectorizedwa ngokuzenzakalela yi-compiler. Ividiyo ekhishiwe iyafana ngokuphelele futhi ijubane likhuphuka izikhathi ezingu-2.7.
Ukunciphisa izindleko zezikhungo zedatha: Ihlaziya yonke idatha ye-telemetry yokusebenza yeqoqo, izitholele futhi isebenzise izixazululo zokuthuthukisa imemori. Ngemva kokukhishwa, ikhulula imemori engaphezu kuka-300 TiB, kanti kulindeleke ukonga okuphelele okuphakathi kuka-500 TiB no-1 PiB.
Ukuthuthukisa ama-algorithm e-quantum: Ilungiselela izinhlelo ezingaphansi ze-quantum computing ngokuthuthukisa umkhiqizo wenani lama-qubit nenani lokusebenza kwama-gate. Esimweni esisodwa, yeqe i-baseline eshicilelwe ngo-40% ngemizuzu embalwa.