I-Google iveze i-Gemini 4 Argon , imodeli yobuhlakani bokwenziwa esezingeni eliphezulu enomgomo ocacile: ukusingatha imisebenzi yobungcweti eyinkimbinkimbi nedla isikhathi. Le nkampani iveze lokhu njengesinyathelo esiphambili sokuhlelwa kwezinhlelo, ubuhlakani bebhizinisi, kanye nokuphepha kwe-inthanethi, yize ukufinyelela kokuqala kungeke kuvulekele umphakathi.
Lesi simemezelo sibeka i-Argon esicongweni samamodeli anamandla kakhulu, kodwa futhi siphakamisa imibuzo eminingana. Akukho usuku oluthile lokukhishwa komphakathi jikelele , futhi imiphumela eminingi yaso ivela ekuhlolweni okushicilelwe yi-Google uqobo, ngakho-ke ukuqapha kusafanele.
Ukukhishwa okulawulwayo okungenalo usuku olubekiwe lomphakathi jikelele
Abokuqala ukuyihlola kuzoba ochwepheshe bezokuphepha kwe-inthanethi, izinhlangano zikahulumeni, kanye nabalingani abathembekile ngaphakathi kohlelo lwe-Fairwind. I-Google ifuna la maqembu ahlole imodeli ngaphambi kokuba amakhono ayo adluliselwe kubabukeli abaningi.
Okulandelayo, imephu yendlela ihlose amakhasimende akhokhelayo e-API kanye nababhalisile be-Google AI Ultra . Akukho usuku oluqinisekisiwe okwamanje kwabanye abasebenzisi, futhi imodeli ayitholakali emphakathini.
Le nkampani iphinde ihlanganyele enkambisweni yokuzithandela lapho uhulumeni wase-US ebuyekeza khona amamodeli athuthukile ngaphambi kokusetshenziswa kwawo. Empeleni, i-Argon ilandela isu elifana nezinye izinguqulo zakamuva: ihlolwa kuqala ezindaweni ezilawulwayo bese ukufinyelela kuyandiswa.
Amathokheni okuphuma afinyelela esigidini
Esinye sezici ezichazayo zale modeli umkhawulo wayo wokukhipha ofika kumathokheni ayisigidi esisodwa , uma kuqhathaniswa nama-64.000 esizukulwaneni esidlule. Ithokheni iyunithi encane kunazo zonke ezisetshenziswa yilezi zinhlelo, ngakho-ke ukuphakamisa lowo mkhawulo kuvumela ukukhiqizwa kwezimpendulo ezinde kakhulu namadokhumenti ngaphandle kokuphazamisa ukuhamba komsebenzi kuzo zonke izinyathelo.
Lolu shintsho luhloselwe imisebenzi edinga ukugxila isikhathi eside: ukubuyekeza ikhodi, ukuhlaziya amadokhumenti, noma ukuqedela izinqubo zezinyathelo eziningi . Nokho, ukuba usizo kwayo kwangempela kuzoncika ekutheni imodeli igcina ikhwalithi njengoba umsebenzi uba mude.
I-Google ithi onjiniyela bayo sebevele basebenzisa i-Argon emisebenzini yansuku zonke, kusukela ekulungiseni amaphutha esofthiwe kuya ekufudukeni kwamakhodi kanye nokuklama i-algorithm. Le nkampani ikhuluma ngokufuduka kusuka ku -C kanye ne-C++ kuya ku-Rust kanye nokulungiswa kwezikhungo zayo zedatha okukhulule inkumbulo engaphezu kuka-300 TiB.
Imiphumela ekubhaleni ikhodi kanye nokuphepha kwe-inthanethi
Ekuhleleni izinhlelo, i-Google ibeka i-Gemini 4 Argon ku- 77,9% ku-DeepSWE v1.1 , ukuhlolwa okwenzelwe imisebenzi yobunjiniyela besofthiwe eqhubeka isikhathi eside kanye nezimo ezifana nomsebenzi wangempela. Le nkampani iveza lokhu njengerekhodi elisha, yize igcizelela ukuthi lokhu kuyidatha yokusebenza hhayi ulwazi oluvaliwe lwabo bonke abasebenzisi.
Ukuphepha kwe-inthanethi kungenye insika ebalulekile. Ku -CWE-bench v1, i-Argon ithola amaphuzu angu-68% futhi ilingana nendawo yokuqala, ngokusho kwe-Google. Imodeli ingathola, iqinisekise, futhi ilungise ubuthakathaka ngezilimi ezingaphezu kwamashumi amabili, futhi ihlaziye ngisho nezinhlelo zewebhu ezibukhoma ngaphandle kokudinga ukufinyelela ikhodi yazo yomthombo.
Le nkampani icaphune ukusebenzisana kwayo neWiz ohlelweni lwe-Scan for Good njengesibonelo. Ngokusho kwalokhu kulandisa, i-Argon ithole ubuthakathaka obukhulu ohlelweni lokunakekelwa kwempilo olusetshenziswa ezibhedlela emazweni amaningana olungadalula idatha yomuntu siqu. Lolu hlelo lubeka phambili izibhedlela, izinsizakalo zomphakathi, kanye nengqalasizinda ebalulekile.
I-Google iphinde yachaza ukuthi i-Argon ifaka ukuvikelwa ekuhlaselweni okusheshayo kokujova kanye nokulawula ukuphepha . Kumodeli enamakhono okuzivikela, leli phuzu libaluleke njengemiphumela yayo yokuhlolwa.
Umsebenzi wobungcweti nokuqonda okubonakalayo
Ngaphandle kwekhodi, i-Google ifuna i-Argon incintisane emisebenzini esekelwe olwazini. Le nkampani ithi ihola i- Vals Index , elinganisa umthelela wezomnotho kwezezimali, izinhlelo, umthetho, kanye nentela, futhi ihamba phambili ekuhlolweni okuthile okufana ne-Vals Finance Agent v2 kanye ne-Harvey's Legal Agent Benchmark.
Ekwenzeni okuzenzakalelayo kwebhizinisi, i-Argon ibekwe endaweni yokuqala ku -AutomationBench kaZapier ngamaphuzu angu-51,3% . Lezi zivivinyo zenzelwe ukukala ukuthi imodeli ingakwazi yini ukwenza imisebenzi yebhizinisi evamile kusukela ekuqaleni kuze kube sekupheleni, hhayi nje ukuphendula imibuzo ehlukanisiwe.
Ukuqonda okubonakalayo nakho kubonakala kakhulu enkulumweni ye-Google. Ku -LVBench, egxile kumavidiyo amade, i-Argon ithola amaphuzu angu-91,7% , okwenza ikwazi ukuhlonza imininingwane ekuqopheni okubanzi futhi yenze izinqumo ngokusekelwe kumadokhumenti noma ihluzo ezahlukahlukene zobungcweti.
I-Google iqhathanisa imodeli ne-GPT-6 Astra kanye ne-Claude Opus 5.5 ematafuleni ayo, ithi i-Argon ihamba phambili ekuhleleni izinhlelo, ezezimali, kanye nomsebenzi wezomthetho. Ngenxa yokuthi le datha ayitholakali kabanzi, amasayithi okuqhathanisa azimele awakwazi ukuphinda lezi zivivinyo , ngakho-ke isimangalo senkampani sokusebenza okuphezulu sisadinga ukuqinisekiswa kwangaphandle.
Intengo kanye nokufinyelela kumabhizinisi nakubasebenzisi
Intengo yonjiniyela izoqala ku- $2 ngesigidi samathokheni atshaliwe kanye no-$10 ngesigidi samathokheni akhishwe ngesikhathi sokwethula. Amathokheni agciniwe azoncishiswa ngo-95%, ngemva kwalokho imali izokwenyuka ibe ngu-$4 ngethokheni ngalinye elitshaliwe kanye no-$20 ngethokheni ngalinye elikhishwe.
ESpain nakulo lonke elaseYurophu, isihluthulelo ukutholakala. I-Argon ayikakwazi ukuhlolwa njengomkhiqizo ovulekile okwamanje; amabhizinisi nochwepheshe kuzodingeka balinde ukufika kwayo nge- API ekhokhelwayo noma ukubhaliswa kwe-Google AI Ultra . I-Google ayikacacisi ukuthi lokho kuzokwenzeka nini.
Lokhu kunyakaza kuhambisana nomkhuba obanzi: amamodeli angaphambili awasaqalwa njalo njenge-chatbot yawo wonke umuntu, kodwa njengamathuluzi aqalwa ukuqaliswa ukuze kukhethwe ozakwethu. Isivinini sokufakwa kanye nokuphepha kubaluleke njengezinombolo zalolu hlobo lwezethulo.
I-Gemini 4 Argon, okwamanje, isalokhu isesigabeni sokufinyelela esinqunyelwe esigxile ekuvikeleni kwe-inthanethi, ohlelweni, kanye nomsebenzi webhizinisi. Amathokheni ayo okuqala ayisigidi, ama-77,9% ku-DeepSWE, kanye nama-68% ku-CWE-bench yizibalo ezethulwa yi-Google, kodwa umthelela wayo wangempela uzobonakala kuphela uma ifinyelela kubasebenzisi abaningi futhi ingalinganiswa ngaphandle kokuhlolwa komkhiqizi.