Nemotron 3.5 Lightning Uyasebenza

I-Nemotron 3.5 Lightning iyimodeli ye-MoE yomthombo ovulekile enamapharamitha angu-30B eyethulwe ngu-NVIDIA, elungiselelwe ngokukhethekile amasistimu anama-agent amaningi. Uma iqhathaniswa namamodeli afanayo, ijubane lokukhipha imiphumela likhuphuka izikhathi ezi-4, kanti ijubane lokuqedela imisebenzi yama-agent likhuphuka ngo-30%, inikeza ibhalansi enhle phakathi kokunemba nejubane. Imodeli isekela ukuthunyelwa kuzo zonke izindawo, kusukela kuma-PC e-RTX, i-Jetson kuya ezikhungweni zedatha, futhi isivele itholakala njengomthombo ovulekile ezinkundleni ezifana ne-Hugging Face ne-NVIDIA Build, okusiza ekwakheni izinhlelo zama-agent ezisebenza kahle kakhulu.

Ukwenza imisebenzi yama-agent amaningi: Le modeli yakhelwe ngokukhethekile imisebenzi ethile kumasistimu amakhulu anama-agent amaningi, isiza abathuthukisi bakhe izinhlelo zama-agent ezihlakaniphe futhi ezisebenza kahle.
Ukukhipha imiphumela ngokushesha: Uma iqhathaniswa namamodeli afanayo, i-Nemotron 3.5 Lightning ingakhuphula ijubane lokukhipha imiphumela kuze kufike ezikhathini ezi-4, isheshise ukuqedwa kwemisebenzi yama-agent ngo-30%.
Ukuthunyelwa okuguquguqukayo kuzo zonke izindawo: Imodeli ingasebenza ngokuqondile kumasistimu e-AI endawo, okuhlanganisa ama-PC e-NVIDIA RTX, i-DGX Spark ne-Jetson, futhi inganwetshwa ngaphandle komthungo iye kumadivayisi asemaphethelweni, eziteshini zokusebenza, ezikhungweni zedatha nasezindaweni zamafu.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI omthombo ovulekile
Isakhiwo se-MoE sabachwepheshe abaxubile: I-Nemotron 3.5 Lightning isebenzisa isakhiwo se-MoE (Mixture of Experts) esinamapharamitha angu-30B, inani eliphelele lamapharamitha lifinyelela ezigidini eziyizinkulungwane ezingu-30, kodwa ekucabangeni ngakunye kuvulwa kuphela inethiwekhi engaphansi enamapharamitha acishe abe yizigidi eziyizinkulungwane ezingu-3. Ngendlela yokuvula okungagcwele, yehlisa kakhulu izindleko zokubala ngenkathi igcina amakhono emodeli enkulu, ngaleyo ndlela inikeze ukucabanga okusebenzayo.
Ukucindezelwa nge-NVFP4: Imodeli isekela ifomethi yokulinganisa i-NVFP4, okuwuhlelo luka-NVIDIA lokunemba kwe-floating-point okungu-4-bit. Lunciphisa kakhulu usayizi wemodeli nokusetshenziswa kwememori yevidiyo, luyenze ikwazi ukusebenza kahle kuma-PC e-RTX, kuma-Jetson nakwamanye amadivayisi asezingeni labathengi nasemaphethelweni.
Ukulungiselelwa okukhethekile kwezimo zama-agent: Imodeli iqeqeshwe futhi yalungiselelwa ngokukhethekile ukubambisana kwama-agent amaningi, igxile ekuthuthukiseni ukuzinza nokunemba kokucabanga okude, ukubiza amathuluzi nokuhlukanisa imisebenzi, ukuze isekele kangcono ukuhamba komsebenzi okuyinkimbinkimbi kwama-agent.
Ukulungiselelwa okuhlangene kwejubane nokunemba: Ekuklanyweni kwesakhiwo nasu lokuqeqesha, i-Nemotron 3.5 Lightning ithuthukisa ngokuhlanganyela ukunemba nejubane lokucabanga. Ekuhlolweni okuyisisekelo kwe-PinchBench, ikhombisa ukusebenza okuphelele okungcono kunamamodeli afanayo, iqinisekisa ikhwalithi yokuqedwa komsebenzi futhi iphakamisa ijubane lonke lokusebenzisa ama-agent ngo-30%.

Ukulanda ku-Hugging Face: Vakashela inqolobane esemthethweni ye-Hugging Face ukuze ulande izisindo zemodeli, bese ulayisha futhi usebenzise i-Nemotron 3.5 Lightning endaweni.
Inkundla ye-NVIDIA Build: Vakashela iwebhusayithi ye-NVIDIA Build ukuze uthole incazelo yemodeli, imibhalo yobuchwepheshe nomhlahlandlela wokuyithumela.
Ukuthunyelwa kumadivayisi endawo: Thumela imodeli ngokuqondile kuma-PC e-NVIDIA RTX, i-DGX Spark, i-DGX Station noma i-Jetson ukuze isebenze kumadivayisi endawo.
Ukwandiswa kwezinga lebhizinisi: Yandisa ngaphandle komthungo iye eziteshini zokusebenza ze-RTX PRO, ezikhungweni zedatha noma ezindaweni zamafu ukuze kuhlangatshezwane nezidingo zezinhlelo zebhizinisi.

Isakhiwo se-MoE sokuvula okungagcwele: Imodeli isebenzisa isakhiwo se-MoE esinamapharamitha angu-30B esewonke kanye namapharamitha acishe abe ngu-3B avulwayo, yehlise kakhulu izindleko zokubala ngenkathi igcina amakhono emodeli enkulu.
Ukucabanga okushesha kakhulu: Uma iqhathaniswa namamodeli afanayo, ijubane lokukhipha imiphumela likhuphuka izikhathi ezi-4, futhi ijubane lonke lokuqedela imisebenzi yama-agent likhuphuka ngo-30%.
Ukusekelwa kokulinganisa i-NVFP4: Ifomethi ka-NVIDIA yokulinganisa i-floating-point engu-4-bit inciphisa kakhulu ukusetshenziswa kwememori yevidiyo, yenze ukuthunyelwa kwenzeke kuzo zonke izindawo, kusukela kuma-PC e-RTX kuya ezikhungweni zedatha.
Ukulungiselelwa okukhethekile kwama-agent: Imodeli iqeqeshelwe ngokukhethekile imisebenzi yama-agent amaningi esebenza isikhathi eside, futhi ithola ibhalansi enhle phakathi kokunemba nejubane lokucabanga.

Iwebhusayithi yephrojekthi：https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightning-delivers-fast-accurate-specialized-task-execution-for-long-running-agents/
Inqolobane yemodeli ye-HuggingFace：https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4

Uhlelo lokubambisana kwama-agent amaningi: Njengemodeli yomsebenzi okhethekile, iqhuba ukuhamba komsebenzi okuyinkimbinkimbi kwama-Agent futhi iqede ngokuphumelelayo ukubizwa kwamathuluzi, ukuhlukaniswa kwemisebenzi nokucabanga okude.
Ukucabanga kwe-AI yasemaphethelweni endaweni: Ngokusebenzisa ukulinganisa kwe-NVFP4 nendlela yokuvula okungagcwele, inikeza ukucabanga kwasendaweni okunokubambezeleka okuphansi nokusetshenziswa kwamandla okuphansi kuma-PC e-RTX nakumadivayisi e-Jetson.
Inkundla yama-agent yezinga lebhizinisi: Inwebeka ngaphandle komthungo iye ezikhungweni zedatha nasezindaweni zamafu, isekele izinsizakalo zama-agent zebhizinisi nezinqubo ezizenzakalelayo ezisebenza kaningi nangesikhathi eside.
Ukuthuthukiswa kwekhodi ne-DevOps ezenzakalelayo: Ikhiqiza, ibuyekeze futhi ilungise ikhodi ngokushesha, isheshise ukwenziwa kwemisebenzi yama-agent ekuthuthukisweni kwesofthiwe nasekuhlanganisweni okuqhubekayo.
Ukuhlaziywa kwedatha ngesikhathi sangempela nokusekelwa kokuthatha izinqumo: Ezimweni zebhizinisi ezidinga impendulo esheshayo, inikeza imiphumela yokucabanga esheshayo nenembile ukuze isize ekuthathweni kwezinqumo ngesikhathi sangempela.