- I-Nemotron 3 lusapho oluvulekileyo lweemodeli, idatha, kunye neelayibrari ezigxile kwi-AI ye-arhente kunye neenkqubo ze-arhente ezininzi.
- Ibandakanya ubukhulu beMoE obuthathu (iNano, iSuper kunye neUltra) kunye noyilo oluhlanganisiweyo kunye noqeqesho olusebenzayo lwe-4-bit kwiNVIDIA Blackwell.
- I-Nemotron 3 Nano ngoku iyafumaneka eYurophu nge-Hugging Face, i-public clouds kunye ne-NIM microservice, enefestile yee-token ezizigidi ezi-1.
- Le nkqubo izaliswe ziidatha ezinkulu, iNeMo Gym, iNeMo RL kunye noMvavanyi wokuqeqesha, ukulungisa kunye nokuphonononga iiarhente ze-AI ezizimeleyo.
Umdyarho wobukrelekrele bokwenziwa usuka kwii-chatbots ezilula nezizimeleyo ukuya kwiinkqubo zeearhente ezisebenzisanayo, ezilawula imisebenzi emide, kwaye kufuneka zihlolwe. Kule meko intsha, I-NVIDIA igqibe ekubeni ithathe inyathelo elicacileyo: ukuvula kungekuphela nje iimodeli, kodwa kunye nedatha kunye nezixhoboukuze iinkampani, ulawulo lukarhulumente kunye namaziko ophando akwazi ukwakha amaqonga awo e-AI ngolawulo oluthe kratya.
Loo ntshukumo iyenzeka I-Nemotron 3, usapho lweemodeli ezivulekileyo ezijolise kwi-AI yeearhente ezininzi Ifuna ukudibanisa ukusebenza okuphezulu, iindleko eziphantsi zokuqikelela, kunye nokungafihlisi. Esi siphakamiso asijoliswanga njengenye nje i-chatbot yenjongo jikelele, kodwa njenge isiseko sokuthumela iiarhente eziqiqa, ezicwangcisa kwaye zenze imisebenzi enzima kumacandelo alawulwayoOku kubaluleke kakhulu eYurophu naseSpain, apho ulawulo lwedatha kunye nokuthobela imithetho kubalulekile.
Usapho oluvulekileyo lweemodeli ze-AI ezizimeleyo nezizimeleyo
I-Nemotron 3 iboniswa njengo inkqubo epheleleyo yendalo: iimodeli, iiseti zedatha, amathala eencwadi, kunye neendlela zokupheka zoqeqesho phantsi kweelayisenisi ezivulekileyo. Ingcinga ye-NVIDIA kukuba imibutho ayisebenzisi nje kuphela i-AI njengenkonzo engacacanga, kodwa inokuhlola okungaphakathi, ilungelelanise iimodeli kwiidomeyini zazo, kwaye izisebenzise kwiziseko zazo, nokuba zikwilifu okanye kumaziko edatha asekuhlaleni.
Le nkampani icwangcisa esi sicwangciso ngokwesibophelelo sayo UMongami AIOorhulumente kunye neenkampani eYurophu, eMzantsi Korea, nakwezinye iindawo bafuna ezinye iindlela ezivulekileyo endaweni yeenkqubo ezivaliweyo okanye zangaphandle, ezidla ngokungahambelani kakuhle nemithetho yabo yokukhusela idatha okanye iimfuno zophicotho-zincwadi. I-Nemotron 3 ijolise ekubeni sisiseko sobugcisa sokwakha iimodeli zesizwe, zecandelo, okanye zenkampani ezinokubonakala nokulawula okukhulu.
Ngaxeshanye, I-NVIDIA yomeleza isikhundla sayo ngaphaya kwezixhobo zekhompyuthaUkuza kuthi ga ngoku, yayiyeyona nto iphambili enika umboneleli weGPU; ngeNemotron 3, ikwayibeka kwindawo yokubonisa nokuqeqesha izixhobo, ikhuphisana ngokuthe ngqo nabadlali abanjengoOpenAI, Google, Anthropic, okanye neMeta, kwaye ichasene neemodeli zeprimiyamu ezifana neNemotron 3. I-SuperGrok inzimaI-Meta iye yanciphisa ukuzibophelela kwayo kwimithombo evulekileyo kwizizukulwana zakutshanje ze-Llama.
Kwinkqubo yophando yaseYurophu kunye nenkqubo yokuqalisa—exhomekeke kakhulu kwiimodeli ezivulekileyo ezigcinwe kumaqonga afana neHugging Face—ukufumaneka kweentsimbi, idatha yokwenziwa, kunye neelayibrari phantsi kweelayisenisi ezivulekileyo kubonisa enye indlela enamandla endaweni iimodeli zamaTshayina kunye namaMerika alawula ukuthandwa kunye nokubekwa kwinqanaba elilinganiselweyo.
Uyilo lwe-MoE oluhlanganisiweyo: ukusebenza kakuhle kweearhente ezinkulu
Uphawu oluphambili lobuchwephesha lweNemotron 3 luphawu Uyilo oluhlanganisiweyo lwengxubevange yeengcali ezifihlakeleyo (MoE)Endaweni yokusebenzisa zonke iiparameter zemodeli kwisigqibo ngasinye, kuphela iqhezu lazo elivuliweyo, iqela leengcali ezibaluleke kakhulu kumsebenzi okanye ithokheni ekubhekiswa kuyo.
Le ndlela ivumela kunciphisa kakhulu iindleko zokubala kunye nokusetyenziswa kwememoriOku kwandisa i-token throughput. Kwii-architectures ze-multi-agents, apho amashumi okanye amakhulu ee-arhente etshintshiselana imiyalezo rhoqo, olu buchule lubalulekile ekuthinteleni inkqubo ekubeni ingaqhubeki kakuhle ngokwexabiso le-GPU kunye nelifu.
Ngokwedatha eyabelwana ngayo yi-NVIDIA kunye neebhentshi ezizimeleyo, iNemotron 3 Nano ifezekisa ukuya kuthi ga kane ngaphezulu kweethokheni ngomzuzwana Xa ithelekiswa neyangaphambili, iNemotron 2 Nano, inciphisa ukuveliswa kweethokheni zokuqiqa ezingeyomfuneko malunga ne-60%. Ngokwenyani, oku kuthetha iimpendulo ezichanekileyo ngokulinganayo okanye nangakumbi, kodwa "zingenamagama" amaninzi kwaye ixabiso liphantsi ngombuzo ngamnye.
Uyilo lwe-hybrid MoE, oludityaniswe neendlela ezithile zoqeqesho, lukhokelele ekubeni Uninzi lweemodeli ezivulekileyo eziphucukileyo zisebenzisa izicwangciso zeengcaliI-Nemotron 3 ijoyina le ndlela, kodwa igxile ngokukodwa kwi-AI ye-arhente: iindlela zangaphakathi ezenzelwe ukulungelelanisa phakathi kwee-arhente, ukusetyenziswa kwezixhobo, ukuphathwa kweemeko ezinde, kunye nocwangciso lwenyathelo ngenyathelo.
Ubungakanani obuthathu: iNano, iSuper, kunye neUltra kwimisebenzi eyahlukeneyo
Usapho lweNemotron 3 luhlelwe ngokwe ubukhulu obuthathu obuphambili bemodeli ye-MoE, zonke zivulekile kwaye zine parameters ezisebenzayo ezincitshisiweyo ngenxa yoyilo lwengcali:
- I-Nemotron 3 Nano: malunga neeparamitha ezingama-30.000 ezigidigidi, malunga Iimpahla ezingama-3.000 ezigidi ngethokheni nganyeYenzelwe imisebenzi ejolise kuyo apho ukusebenza kakuhle kubalulekile: ukulungisa iimpazamo zesoftware, ukushwankathela amaxwebhu, ukufumana ulwazi, ukubeka esweni inkqubo, okanye abancedisi abakhethekileyo be-AI.
- I-Nemotron 3 Super: malunga neeparamitha ezili-100.000 leebhiliyoni, kunye Iimpahla ze-10.000 yezigidigidi kuyo yonke inyathelo. Ijoliswe Ukuqiqa okuphucukileyo kwizakhiwo ze-arhente ezininzinge-latency ephantsi nangona iiarhente ezininzi zisebenzisana ukusombulula ukuhamba okuntsonkothileyo.
- I-Nemotron 3 Ultra: umgangatho ophezulu, oneeparamitha ezimalunga nama-500.000 eebhiliyoni ukuya kuthi ga Iimpahla ezingama-50.000 ezigidi ngethokheni nganyeIsebenza njengenjini enamandla yokuqiqa yophando, ucwangciso lwesicwangciso, inkxaso yezigqibo ezikumgangatho ophezulu, kunye neenkqubo ze-AI ezifuna kakhulu.
Enyanisweni, oku kuvumela imibutho Khetha ubungakanani bemodeli ngokwebhajethi yakho kunye neemfuno zakhoI-Nano yemithwalo enzima kakhulu kunye neendleko ezixineneyo; Ingcono kakhulu xa kufuneka ubunzulu bokuqiqa kunye neearhente ezininzi ezisebenzisanayo; kunye ne-Ultra kwiimeko apho umgangatho kunye nomxholo omde udlula iindleko ze-GPU.
Okwangoku YiNemotron 3 Nano kuphela ekhoyo ukuze isetyenziswe ngoko nangoko.Iindidi zeSuper kunye neUltra zicwangciselwe isiqingatha sokuqala sowama-2026, zinika iinkampani zaseYurophu kunye neelabhoratri ixesha lokuzama kuqala iNano, ukuseka imibhobho, kwaye kamva, ukufudusa amatyala afuna amandla amakhulu.
I-Nemotron 3 Nano: ifestile yethokheni yezigidi ezili-1 kunye nexabiso eliqulathiweyo
I-Nemotron 3 Nano, njenganamhlanje, iyasebenza. intloko esebenzayo yosaphoI-NVIDIA iyichaza njengeyona modeli ingabizi kakhulu ngokwezibalo kuluhlu, elungiselelwe ukubonelela ngentsebenzo ephezulu kwimisebenzi ye-arhente ezininzi kunye nemisebenzi enzima kodwa ephindaphindwayo.
Phakathi kweempawu zayo zobugcisa, ezi zilandelayo zibalaseleyo: ifestile yomxholo weethokheni ezifikelela kwisigidi esinyeOku kuvumela ukugcinwa kwenkumbulo yamaxwebhu amaninzi, iindawo zokugcina iikhowudi ezipheleleyo, okanye iinkqubo zoshishino ezinamanyathelo amaninzi. Kwizicelo zaseYurophu kwiibhanki, ezempilo, okanye kulawulo lukarhulumente, apho iirekhodi zinokuba zininzi, olu lwazi lubaluleke kakhulu.
Iimpawu zombutho ozimeleyo Uhlalutyo lokwenziwa lubeka iNemotron 3 Nano njengenye yeemodeli ezivulelekileyo kakhulu Idibanisa ubukrelekrele, ukuchaneka, kunye nesantya, kunye namazinga okukhupha kwi-token ezingamakhulu ngomzuzwana. Olu dibaniso lwenza ukuba ibe nomtsalane kubahlanganisi be-AI kunye nababoneleli beenkonzo eSpain abafuna amava amahle omsebenzisi ngaphandle kweendleko eziphezulu zeziseko.
Ngokuphathelele iimeko zokusetyenziswa, i-NVIDIA ijolise kwiNano kwi Isishwankathelo somxholo, ukulungisa iimpazamo zesoftware, ukufunyanwa kolwazi, kunye nabancedisi be-AI yeshishiniNgenxa yokuncipha kweethokheni zokuqiqa ezingafunekiyo, kunokwenzeka ukuba kuqhutywe iiarhente ezigcina incoko ende nabasebenzisi okanye iinkqubo ngaphandle kokuba ityala le-inference linyuke kakhulu.
Idatha evulekileyo kunye neelayibrari: iNeMo Gym, iNeMo RL kunye noMvavanyi
Enye yezona zinto zahlukileyo kwiNemotron 3 kukuba Akupheleli nje ekukhululeni ubunzima bemodeliI-NVIDIA ihamba nosapho kunye neqela elibanzi lezixhobo ezivulelekileyo zokuqeqesha, ukulungisa, kunye nokuvavanya iiarhente.
Kwelinye icala, yenza ukuba kufumaneke i-corpus yokwenziwa iitoni ezininzi zezigidigidi zedatha yangaphambi koqeqesho, emva koqeqesho, kunye nokuqinisaEzi seti zedatha, ezigxile ekuqiqeni, ekubhaleni iikhowudi, nakwimisebenzi yamanyathelo amaninzi, zivumela iinkampani kunye namaziko ophando ukuba avelise iintlobo zazo zeNemotron (umz., ezomthetho, ezempilo, okanye ezemizi-mveliso) ngaphandle kokuqala ekuqaleni.
Phakathi kwezi zixhobo, ezi zilandelayo zivelele: Iseti yedatha yoKhuseleko lwe-Nemotron AgenticIqokelela idatha ye-telemetry malunga nokuziphatha kwe-arhente kwiimeko zokwenyani. Injongo yayo kukunceda amaqela ukuba alinganise kwaye aqinise ukhuseleko lweenkqubo ezizimeleyo ezintsonkothileyo: ukusuka kwizenzo ezenziwa yi-arhente xa idibana nedatha ebucayi, ukuya kwindlela esabela ngayo kwimiyalelo engacacanga okanye enokuba yingozi.
Ngokuphathelele icandelo lezixhobo, i-NVIDIA iyaqalisa I-NeMo Gym kunye ne-NeMo RL njengeelayibrari ezivulelekileyo yoqeqesho lokuqinisa kunye noqeqesho lwasemva koqeqesho, kunye noMvavanyi weNeMo wokuvavanya ukhuseleko kunye nokusebenza. La mathala eencwadi abonelela ngeendawo zokulinganisa ezilungele ukusetyenziswa kunye nemibhobho kunye nosapho lweNemotron, kodwa inokwandiswa nakwezinye iimodeli.
Zonke ezi zinto—iintsimbi, iiseti zedatha, kunye nekhowudi—zisasazwa I-GitHub kunye ne-Hugging Face zinelayisenisi phantsi kwe-NVIDIA Open Model License.ukuze amaqela aseYurophu akwazi ukuyidibanisa ngaphandle komthungo kwii-MLOp zawo. Iinkampani ezifana nePrime Intellect kunye ne-Unsloth sele zifake i-NeMo Gym ngqo kwimisebenzi yazo ukuze kube lula ukufunda ngokuqiniswa kwiNemotron.
Ukufumaneka kwiilifu zikawonke-wonke kunye nenkqubo yendalo yaseYurophu
I-Nemotron 3 Nano ngoku iyafumaneka apha Ukujongana nobuso y GitHubkunye nababoneleli beenkcukacha ezifana neBaseten, DeepInfra, Fireworks, FriendliAI, OpenRouter, kunye neTogether AI. Oku kuvula ithuba kumaqela ophuhliso eSpain lokuvavanya imodeli nge-API okanye ukuyithumela kwiziseko zayo zophuhliso ngaphandle kobunzima obukhulu.
Kwicala elingaphambili lelifu, INemotron 3 Nano ijoyina i-AWS nge-Amazon Bedrock ngenxa yokuqonda okungenaseva, kwaye ibhengeze inkxaso kwiGoogle Cloud, CoreWeave, Crusoe, Microsoft Foundry, Nebius, Nscale, kunye neYotta. Kwimibutho yaseYurophu esele isebenza kwezi qonga, oku kwenza kube lula ukusebenzisa iNemotron ngaphandle kotshintsho olukhulu kulwakhiwo lwayo.
Ukongeza kwilifu loluntu, iNVIDIA ikhuthaza ukusetyenziswa kweNemotron 3 Nano njenge Inkonzo encinci ye-NIM ingasetyenziswa nakweyiphi na isiseko esikhawulezileyo se-NVIDIAOku kuvumela iimeko ezixutyiweyo: inxalenye yomthwalo kumafu aphesheya kunye nenxalenye kumaziko edatha asekuhlaleni okanye kumafu aseYurophu abeka phambili ukuhlala kwedatha kwi-EU.
Iinguqulelo I-Nemotron 3 Super kunye ne-Ultra, ezijolise kwimisebenzi enzima yokuqiqa kunye neenkqubo ezinkulu ze-multi-agents, kucwangciselwe isiqingatha sokuqala sika-2026Eli xesha livumela uphando lwaseYurophu kunye nenkqubo yezoshishino ukuba yenze uvavanyo ngeNano, iqinisekise iimeko zokusetyenziswa, kwaye iyile amaqhinga okufudukela kwiimodeli ezinkulu xa kuyimfuneko.
I-Nemotron 3 ibeka iNVIDIA njengomnye wababoneleli abaphambili be- iimodeli ezivulekileyo eziphezulu ezilungiselelwe i-AI ye-arhenteNgesiphakamiso esidibanisa ukusebenza kakuhle kobuchwephesha (i-hybrid MoE, i-NVFP4, umxholo omkhulu), ukuvuleleka (iisisindo, iiseti zedatha kunye neelayibrari ezikhoyo) kunye nokugxila ngokucacileyo kulawulo lwedatha kunye nokungafihlisi, iinkalo ezibaluleke kakhulu eSpain nakwamanye amazwe aseYurophu, apho imithetho kunye noxinzelelo lokuhlola i-AI zikhula ngakumbi.
Ndingumntu othanda itekhnoloji ojike umdla wakhe we "geek" waba ngumsebenzi. Ndichithe ngaphezulu kweminyaka eli-10 yobomi bam ndisebenzisa itekhnoloji yokusika kwaye ndikhenkceza ngazo zonke iintlobo zeenkqubo ngenxa yokufuna ukwazi okumsulwa. Ngoku ndiqeqeshelwe ubugcisa bekhompyutha nakwimidlalo yevidiyo. Oku kungenxa yokuba ngaphezu kweminyaka emi-5 ndibhalela iiwebhusayithi ezahlukeneyo kwitekhnoloji kunye nemidlalo yevidiyo, ndisenza amanqaku afuna ukukunika ulwazi oludingayo ngolwimi oluqondakalayo kuye wonke umntu.
Ukuba unayo nayiphi na imibuzo, ulwazi lwam lusuka kuyo yonke into enxulumene nenkqubo yokusebenza yeWindows kunye ne-Android yeefowuni eziphathwayo. Kwaye ukuzinikela kwam kukuwe, ndihlala ndikulungele ukuchitha imizuzu embalwa kwaye ndikuncede usombulule nayiphi na imibuzo onokuba nayo kweli lizwe le-intanethi.