I-Google ithathe isinyathelo esibalulekile esu layo ubuhlakani bokwenziwa obuvulekile Ngokuqaliswa kwe-Gemma 4, umndeni omusha wamamodeli uhlose ukuhlanganisa amazinga aphezulu okucabanga nezidingo zehadiwe eziphansi kakhulu. Le nkampani iveza lesi sizukulwane njengenye indlela ebaluleke kakhulu kulabo abadinga ukusebenzisa i-AI ethuthukisiwe kwingqalasizinda yabo, kusukela kumadivayisi eselula kuya ezikhungweni zedatha.
Kunokuba yimodeli eyodwa, I-Gemma 4 iwuhlu oluphelele lwezinhlobo ezine ezivulekileYakhelwe onjiniyela, amabhizinisi, kanye nezinhlangano zomphakathi ezifuna ukulawula okwengeziwe kwedatha yazo kanye nokusetshenziswa kwayo, lesi siphakamiso sifanelana kahle kakhulu nezidingo ze Ubukhosi bedijithali kanye nokuthobela imithetho eYurophulapho amandla okunquma ukuthi ukusetshenziswa kuphi nokuthi idatha igcinwa kuphi eba yinto ebaluleke kakhulu.
Umndeni wamamodeli amane ogxile "ekuhlakanipheni okusekelwe kupharamitha"
I-Gemma 4 yakhiwe phezu kwesisekelo sobuchwepheshe esifanayo nesakho IGemini 3Kodwa ngenhloso ecacile: ukukhulisa lokho i-Google ekubizayo "ubuhlakani obusekelwe kupharamitha"Esikhundleni sokuncintisana ngobukhulu kuphela, inkampani izigqaja ngokufinyelela amazinga okusebenza afana nezinhlelo ezinkulu kakhulu kumamodeli amancane.
Umndeni uqukethe osayizi abane abahlukene: Isebenza kahle 2B (E2B), Isebenza kahle 4B (E4B), imodeli ye 26B enomklamo we-Mixture of Experts (MoE) kanye nohlobo oluqinile lwe 31B amapharamithaLesi sakamuva sesivele sitholakala ku- abaphezulu abathathu kuzinga le-Arena AI lamamodeli avulekile, idlula ezinye izindlela eziphindaphinda inani layo lamapharamitha ngamashumi amabili, into ebaluleke kakhulu kulabo abafuna ukunciphisa izindleko ze-GPU ngaphandle kokulahlekelwa ikhwalithi.
Imodeli ye 26B MoE Yenzelwe ukuthi isebenze ingxenye encane kuphela yamapharamitha ayo (cishe ama-3,8B) ekucabangeni, ithuthukisa isivinini sokukhiqiza amathokheni kanye nokusebenza kahle kwamandla. Ngokuphambene nalokho, inguqulo 31B obukhulu Ibekwe njengenketho ekhethwayo yokulungisa imisebenzi enzima, ukuhlela okuyinkimbinkimbi, kanye nokusetshenziswa okujulile ezindaweni zebhizinisi noma zezikhungo.
I-Google igcizelela ukuthi, ngokwezilinganiso zomphakathi, lezi zinhlobo zincintisana ngqo namamodeli asindayo avela kwabanye abahlinzeki, okuhlanganisa lawo avela kubakhiqizi baseShayina njenge-DeepSeek noma i-Qwen, eminyakeni yamuva nje eye yaba namandla ohlelweni lwe-open source. I-31B kaGemma 4 ibhalwe njengemodeli yesithathu evulekile ehamba phambili ku-Arena AI, kuyilapho i-26B MoE nayo isezingeni eliphezulu.
Ngokombono webhizinisi, lobo budlelwano phakathi kobukhulu nokusebenza busho ukuthi ukusetshenziswa kwemali okuncane kwehadiwe, ukubambezeleka okuphansi kanye nekhono lokusebenzisa amamodeli emingcele ngendlela eyodwa I-NVIDIA H100 80GB GPULokhu kuvula ithuba lokuthi izinkampani zaseYurophu eziphakathi nendawo zisebenze ne-AI ethuthukisiwe ngaphandle kokutshala imali engqalasizinda engalingani.
I-AI encane: iselula, i-IoT kanye nekhompyutha enqenqemeni
Amamodeli amancane, I-E2B kanye ne-E4BZenzelwe ngokuqondile ukusebenza emaphethelweni enethiwekhi, okungukuthi, ku amadivayisi eselula, i-IoT kanye nehadiwe yendawoI-Google iphawula ukuthi lezi zinhlobo zenzelwe ukusebenza Ama-smartphone e-Android, Ukusajingijolo PiJetson Nano kanye nezinye izinhlelo zamandla aphansi, ezine-latency ephansi kakhulu futhi ngisho nangaphandle kokuxhumeka kwe-inthanethi.
Kulesi sigaba, okubaluleke kakhulu akusikho nje amandla aluhlaza, kodwa nekhono lokunikeza imisebenzi eminingi kanye nokuphendula okusheshayo ezindaweni ezincishisiwe izinsiza. Amamodeli e-Gemma 4 onqenqemeni angakwazi ukusingatha umbhalo, izithombe namavidiyofuthi uma kwenzeka i-E2B kanye ne-E4B zengeza ukwesekwa kwendabuko kwe- alalelwayoLokhu kuvumela izimo zokusetshenziswa ezifana nabasizi bezwi bendawo, ukuqashelwa kwesithombe sasensimini, noma ukuhlaziywa kwevidiyo ngesikhathi sangempela ngaphandle kwesidingo sokuthumela idatha efwini.
Ifasitela lomongo lala mamodeli alula lifinyelela ku- Amathokheni angama-128.000Lokhu kwanele ukucubungula amadokhumenti amade, izingxoxo ezinde, noma izingcezu zekhodi ezifanele ngesaziso esisodwa. Ngokusho kwe-Google, lokhu kuhlanganiswa komongo obanzi kanye nokusebenza kwendawo kusiza ukuqeda ukungezwani. ubumfihlo, ukuxhumana kanye nokubambezelekaLokhu kubaluleke kakhulu kumaphrojekthi ezimboni, ezempilo, noma ezemfundo eYurophu, lapho imikhawulo ekucutshungulweni kwedatha iba lukhuni kakhulu.
Ngokombono wabakhiqizi behadiwe, i-Gemma 4 ivula umnyango wokuhlanganiswa I-AI ethuthukisiwe ngqo emikhiqizweni yabathengiKusukela kuma-smartphone namathebulethi kuya kumadivayisi ezokwelapha kanye nezinzwa zezimboni, inkampani igcizelele ukuthi la mamodeli aklanyelwe ukusebenza nama-chip avela kubahlinzeki bezinhlelo zemvelo ze-Android abavamile, njenge-Qualcomm kanye ne-MediaTek, okwenza kube lula ukuwasebenzisa kabanzi.
Ngaphezu kwalokho, ukwakheka kwamamodeli asemaphethelweni kusebenzisa amasu afana nalawa: Ukushumeka Ngesendlalelo Ngasinye (PLE) ukukhulisa ukusebenza kahle kokusetshenziswa kwamapharamitha, okuvumela ukucabanga nokuqonda umongo ngezindleko eziphansi kakhulu zokubala kunokujwayelekile kumamodeli enhloso ejwayelekile.
I-Multimodality, ama-ejenti, kanye nosekelo lonjiniyela oluthuthukisiwe
Enye yamandla e-Gemma 4 ukuzibophezela kwayo okucacile ku- ukugeleza komsebenzi we-agentAmamodeli awagcini nje ngokukhiqiza umbhalo: ahlanganiswa ngokomdabu ukubizwa komsebenzi, umphumela we-JSON ohleliwe, kanye nemiyalelo yesistimuLokhu kuvumela ukwakhiwa kwama-ejenti azimele ahlela izinyathelo ezahlukahlukene, abize ama-API angaphandle, futhi abuyisele imiphumela kumafomethi ahlanganiswe kalula nezinhlelo zokusebenza zebhizinisi.
I-Google igcizelela ukuthi wonke amamodeli emndenini we-Gemma 4 aklanywe njenge- izizathu eziphezulungezindlela zokucabanga ezilungisekayo ukuze kulungiswe ukujula kokucabanga ngokwemisebenzi. Lokhu kuhumushela emiphumeleni engcono ku ukucabanga okunezigaba eziningi, ukukhiqizwa kwekhodi okungaxhunyiwe ku-inthanethi kanye nokuxazulula izinkinga okuyinkimbinkimbi, izici ezibalulekile ezindaweni zokuphatha izinkampani kanye nomphakathi lapho kudingeka khona ukwethembeka.
Endizeni ye-multimodal, amamodeli amane angacutshungulwa Umbhalo nezithombe ezinezinqumo ezahlukene kanye nezilinganiso zokubukeka, kuyilapho izinhlobo ze-E2B kanye ne-E4B zandisa lowo mthamo ukuze ividiyo nomsindoLokhu kuhlanganiswa kwenza kube nokwenzeka, isibonelo, izinhlelo ezihlaziya amadokhumenti ngezithombe, amavidiyo okuqapha ezimboni, noma okuqukethwe okucebile kwezemfundo, futhi zikhiqize izimpendulo zomongo ngesikhathi sangempela.
Ifasitela lomongo lifinyelela ku- Amathokheni angu-256.000 kumamodeli amakhulu kakhuluLokhu kuvumela abasebenzisi ukuthi balayishe wonke amakhodi agciniwe, izinkontileka zomthetho ezinde, noma imibhalo eminingi yobuchwepheshe embuzweni owodwa. Ukuze uthole ukwesekwa, ukubonisana, noma amaqembu okuhlola i-IT, lokhu kwenza kube lula ukwenza imisebenzi ezenzakalelayo eyayidinga amahora amaningi okubuyekezwa ngesandla ngaphambilini.
Ngokuphathelene nezilimi, i-Gemma 4 isekela okungaphezu kwalokho Izilimi ezingama-140EYurophu, futhi ikakhulukazi eSpain, lokhu kusho ukuthi izixazululo zezilimi eziningi zingathuthukiswa ezihlanganisa konke kusukela ezilimini eziyinhloko ze-EU kuya ezilimini ezingamelwe kakhulu, okusiza ekuhlangabezaneni nemigomo yokufinyeleleka nokufakwa ezinsizeni zomphakathi nezangasese.
Ukuhlanganiswa kwamafu, ubukhosi bedijithali, kanye nokuthunyelwa eYurophu
Ukufakwa kwe-Gemma 4 akugcini nje ngehadiwe yasendaweni. I-Google ihlanganise lawa mamodeli ekunikezelweni kwayo kwamafu ngokusebenzisa IVertex AI y I-Google Kubernetes Engine (GKE)okuvumela izinhlangano ukuthi zilungiselele izinsizakusebenza zokubala ezizinikele futhi zilinganise imithwalo yemisebenzi yokucabanga ngokufunwa. Kwemikhakha yaseYurophu elawulwayo, lokhu kuhlanganiswe nezinketho ze Ifu Elibusayo kanye nokufakwa kwezikhala emoyeni noma endaweni, okulungisiwe ngokwezidingo zokuhlala kwedatha kanye nokuhambisana ne-General Data Protection Regulation (GDPR).
Le nkampani igcizelela ukuthi izisindo zokunemba ze-bfloat16 zamamodeli amakhulu zingasebenza kahle ku- i-GPU eyodwa engu-80GB NVIDIA H100ukunciphisa isithiyo sokungena kwezinkampani eziphakathi nendawo noma izikhungo zomphakathi ezifuna ukugcina ukulawula ingqalasizinda yazo. Ezinguqulweni ezilinganiselwe, amamodeli angasebenza futhi ihadiwe yabathengi noma izindawo zokusebenza, ukwandisa ububanzi bokusetshenziswa okungenzeka.
Kubaphathi bezobuchwepheshe eSpain nakwamanye amazwe aseYurophu, le nhlanganisela ye imodeli evulekile, ukuthunyelwa okulawulwayo, kanye nokusekelwa kwefu elizimele Ivumela ukwakheka kwezakhiwo ezihlanganisiwe: ingxenye yobuhlakani ingahlala ezikhungweni zedatha zasendaweni, kuyilapho eminye imisebenzi engazweli kakhulu isebenza efwini lomphakathi, konke lokhu kugcina isisekelo sobuchwepheshe esifanayo.
Ngaphezu kwalokho, i-Google inikeza Ikhithi yokuthuthukisa yomenzeli (ADK)Uhlaka oluyi-modular olwenza kube lula ukudala, ukuhlola, kanye nokusebenzisa ama-ejenti asekelwe ku-Gemma 4. Luphinde luthembele ezinsizakalweni ezifana nalezi Ukusebenza Kwamafu nge-NVIDIA RTX PRO 6000 GPUs (Blackwell) kwimodi engenaseva, evumela ukuthi amaphrojekthi okuhlola anamandla aphezulu aqaliswe ngaphandle kwesidingo sokuthola ihadiwe yakho kusukela ngosuku lokuqala.
Esimweni saseYurophu lapho impikiswano nge-AI ivame ukugxila ekulawuleni, ekubonakaleni, kanye nasekuhlolweni, kungenzeka ukuthi Sebenzisa amamodeli avulekile ngaphansi kwe-Apache 2.0 kuzakhiwo ezilawulwayo Kukhanga kakhulu abaphathi, amabhange, izinkampani zomshuwalense noma izinkampani emkhakheni wezempilo ezidinga ukuvumelanisa ukusungula izinto ezintsha nezinhlaka zomthetho eziqinile.
Ilayisensi ye-Apache 2.0, i-ecosystem evulekile kanye nokuthambekela komphakathi
Uma kukhona isici esisodwa esidale isithakazelo esithile emphakathini, yisinqumo sokunikeza ilayisense i-Gemma 4 ngaphansi I-Apache 2.0Izinguqulo zangaphambilini ze-Gemma zazisebenzisa amalayisense enziwe ngokwezifiso aphakamisa imibuzo yezomthetho ngemikhiqizo yezentengiselwano; manje, ngelayisensi ejwayelekile yomthombo ovulekile, Onjiniyela nezinkampani bangashintsha, basabalalise kabusha, futhi benze imali ngamamodeli ngokuncintisana okuncane kakhulu.
Lokhu kuvulwa kufike ngesikhathi lapho i-Google izama khona thola indawo kabusha ohlelweni lwe-ecosystem lwamamodeli avulekileLokhu kuza ngemva kwesikhathi lapho ezinye izindlela ezifana ne-Meta's Llama noma amamodeli aseShayina (DeepSeek, Qwen, GLM, Minimax) athole amazinga okwamukelwa. Amazwi anethonya kulo mkhakha, njengomsunguli we-Hugging Face, achaze lesi sinyathelo "njengesinyathelo esikhulu" se-AI yendawo, egcizelela ukuthi amaqembu ezomthetho manje anohlaka olucacile kakhulu lokuvuma amaphrojekthi asekelwe ku-Gemma 4.
Indawo ezungezile umndeni wakwaGemma yayivele ikhombisa amandla ngaphambi kwalolu hlobo. I-Google iphawula ukuthi izizukulwane zangaphambilini zidlula... Ukulandwa kwezigidi ze-400 nokuthi umphakathi udale okungaphezu kwalokho 100.000 ukwehluka ezivumelaniswe nezilimi ezahlukene kanye nezindlela zokusebenzisa. Phakathi kwezibonelo eziphawuleka kakhulu kukhona amamodeli akhethekile IsiBulgaria noma amathuluzi okucwaninga ngomdlavuza afana nalawa Isilinganiso Semisho Seseli 2 yathuthukiswa eNyuvesi yaseYale.
Nge-Gemma 4, inkampani inethemba lokuthi i-"Gemmaverse" izokhula kakhulu, imeme... Ama-startups aseYurophu, amanyuvesi kanye nezikhungo zocwaningo ukuze bakhe izakhi zabo. Ukuhlanganiswa kwelayisensi evunyelwe kanye nezinsimbi ezivulekile kuvumela ukuthuthukiswa kwezinguqulo ezigxile emikhakheni ethile, njengokunakekelwa kwempilo, ubulungiswa, i-Industry 4.0, noma imfundo, okungabe sekwabelwana noma kukhangiswe ngaphandle kwemingcele eminingi kakhulu.
Ezinkampanini zaseSpain, lesi simo sisho ukuthi kungenzeka ukwakha izixazululo ezizimele ku-Gemma 4—njengabasizi bangaphakathi, izinjini zokusesha zezinkampani, noma izinhlelo zokuhlaziya ezithuthukisiwe—ngenkathi kugcinwa ukulawula ikhodi, idatha, kanye nengqalasizinda, into ehambisana kahle nomkhuba wokuqinisa Ubukhosi bezobuchwepheshe baseYurophu.
Amacala okusetshenziswa: kusukela ezinhlanganweni ezintsha kuya ezinkampanini ezinkulu
I-Gemma 4 inikezwe izinhlobo eziningi ze- izicelo ezingaba khonaEzweni lebhizinisi, amamodeli angasetshenziswa ukudala abasizi abasebenzisa izilimi eziningi abasebenzisa i-inthanethi okwaziyo ukusingatha imibuzo eyinkimbinkimbi ngokusebenzisa ukucabanga okuthuthukisiwe, noma ukwenza ngokuzenzakalelayo ukwenziwa kwekhodi nokubuyekezwa kwamaqembu okuthuthukisa.
Amamodeli amakhulu aqondiswe emisebenzini efana ukuhlelwa kwama-ejenti, ukuhlaziywa kwemibhalo eminingiLokhu kufaka phakathi ukukhiqiza imibiko yobuchwepheshe noma ukusiza iminyango yezomthetho kanye neyokuthobela imithetho. Ukuhlanganiswa kwamafasitela omongo obanzi kanye nokusekelwa kwezindlela eziningi kwenza kube lula nge-ejenti eyodwa ukusebenza ngezinkontileka, ama-imeyili, amashadi, izithombe zesistimu yokuqapha, kanye nokuqoshwa komsindo, konke ngaphakathi komsebenzi ofanayo.
Kwezemfundo kanye nasemkhakheni kahulumeni, ikhono lokucubungula umbhalo, izithombe, futhi kwezinye izimo ividiyo nomsindo, kuvumela ukudalwa amapulatifomu okusekela ukufunda ezikhiqiza izifinyezo, izincazelo zesinyathelo ngesinyathelo, noma izinto ezilungiselelwe amazinga ahlukene. Ukusetshenziswa kwendawo kusiza nokuhlonipha izidingo zobumfihlo lapho kusetshenzwa ngedatha ebucayi yabantwana noma amaqembu asengozini.
Enkundleni yabaqalayo, i-Gemma 4 ingaba yisisekelo se imikhiqizo eqondile Ku-fintech, impilo yedijithali, ezokuthutha, noma i-B2B SaaS, ngenxa yokuguquguquka okunikezwa yi-Apache 2.0, amaqembu angenza ukuhleleka Kwayo yemodeli kudatha yabo, bayithumele endaweni noma efwini bese bethengisa umphumela ngaphandle kokuboshwa ngamalayisense aqinile obunikazi.
Okuthakazelisa kakhulu eYurophu ukuthi kungenzeka ukuthi kuthuthukiswe izixazululo ze-AI zasendaweni ezihlonipha imithetho kazwelonke neyomphakathi, isibonelo, ngokugcina idatha ezikhungweni zedatha ezitholakala endaweni yaseYurophu nokugcina amamodeli ngaphansi kokulawulwa okuqondile kwenhlangano, okungaba yisihluthulelo samaphrojekthi ahlobene noMthethonqubo we-EU AI wesikhathi esizayo.
Ungayifinyelela kuphi futhi kanjani i-Gemma 4
I-Google yenze ukuthi izisindo ze-Gemma 4 zitholakale ngeziteshi ezahlukahlukene ukuze kube lula ukwamukelwa kwazo ngabathuthukisi nabacwaningi. Izisindo ezivulekile zingalandwa ku- Ubuso Obumbambayo y GitHub, kuyilapho ukusetshenziswa nge-interface nama-API kutholakala ku- I-Google AI StudioUkuhlanganiswa nakho kunikezwa U-OllamaI-Docker, i-Kaggle, namathuluzi afana ne-LM Studio.
Ngokusho kwenkampani, i-Gemma 4 ingaqhutshwa endaweni yangakini ku-«izigidigidi zamadivayisi e-Android» nakuzo zonke izinhlobo zehadiwe: kusukela Ama-GPU e-Laptop kanye nezindawo zokusebenza, kuze kufike kuma-accelerator onjiniyela azinikele. Lokhu kuhambisana nesu lokwelula i-AI ethuthukisiwe ngale kwezikhungo ezinkulu zedatha, kumadivayisi abasebenzisi bokugcina kanye nezimo zokubala ezisemaphethelweni.
Kulabo abafuna ukuqala ngezivivinyo ezisheshayo, inketho eqondile kakhulu ukusebenzisa I-Google AI Studio kumamodeli angu-26B kanye no-31B noma Google AI Edge Gallery uma kuziwa ezinhlotsheni ze-E2B kanye ne-E4B. Ngesikhathi esifanayo, imiphakathi yabathuthukisi ezinkundleni ezifana ne-Hugging Face isivele ishicilela izinguquko kanye nokulungiselelwa okulungele ukusetshenziswa kwezindawo ezahlukene.
ESpain nakwamanye amazwe aseYurophu, kulindeleke ukuthi abahlanganisi bendawo kanye nabahlinzeki bezinsizakalo abaphethwe bazoqala ukunikeza izixazululo ze-turnkey ezisekelwe ku-Gemma 4, ukuhlanganisa ukuthunyelwa kwamafu okuzimela, ukwesekwa ngesiSpanishi kanye nokuzivumelanisa nemithethonqubo ethile yomkhakha, njengaleyo yezinsizakalo zezezimali noma ukunakekelwa kwempilo.
Sekukonke, ukwethulwa kwe-Gemma 4 kubeka i-Google njengomunye wabadlali abafanele kakhulu emkhakheni we- amamodeli e-AI avulekile nasebenzisekayo endaweni, ngesikhathi lapho imboni yaseYurophu ifuna khona ngqo amathuluzi ahlanganisa ukusebenza okuphezulu, ukulawula idatha kanye nezinhlaka ezicacile zamalayisense ukwakha imikhiqizo yezentengiselwano yesikhathi eside.
