I-API ye-Nemotron Engavinjelwe

Iningi lamamodeli "avulekile" livulekile ngezisindo kuphela futhi akukho okunye — isigaxa samapharamitha, ilayisensi, kanye nokuthula mayelana nokuthi lezi zinombolo zavelaphi. I-Nemotron ibilokhu iyisichibi esiqhubekayo iminyaka emibili: i-NVIDIA ithumela izisindo, amaqoqo edatha okuqeqeshwa kwangemuva, izindawo ze-RL kanye nemaresiphi. Lokho kubalulekile esiphethweni esingavinjelwe ngendlela okulula ukuyeqa. Imodeli enedatha yokuqondaniswa eshicilelwe yimodeli onokucabanga ngokuziphatha kwayo kokwenqaba esikhundleni sokuqagela — futhi eyimodeli inkampani yesithathu engayikhonza ngokusemthethweni ngaphandle kwesihlungi sokuqukethwe esibotshelwe emlonyeni wayo.

01Iyini i-Nemotron 3 Ultra?

I-Nemotron 3 Ultra yilungu elikhulu kunawo wonke omndeni we-Nemotron 3 we-NVIDIA, elakhishelwa ku-Hugging Face ngo-Nhlangulana 4, 2026. Umbiko wobuchwepheshe we-NVIDIA uyichaza njenge-"550 billion total and 55 billion active parameter Mixture-of-Experts Hybrid Mamba-Attention language model" — cishe u-10% wenethiwekhi usebenza kunoma yiliphi ithokheni, okuyikho okwenza imodeli yamapharamitha ayisigamu sesigidintathu ikwazi ukukhonzwa nhlobo.

Ukwakheka okubikiwe kungamangqimba angu-108 kubukhulu bemodeli obungu-8,192, kuxutshwa amangqimba e-Mamba-2 e-state-space nesethi esabalele yamangqimba okunaka aphelele, ochwepheshe abangu-512 engqimbeni ngalinye le-MoE ngokuqondiswa kwe-top-22. Ukuqeqeshwa kwangaphambili kwaqhutshwa cishe emathokhenini ombhalo ayizigidigidi ezingu-20 ngezigaba ezimbili — okuqale kwathambekela ekuhlukahlukeni, kwabe sekuthambekela ekhwalithi — ngomncamo ka-Septhemba 2025, izilimi zemvelo eziyi-12 nezilimi zezinhlelo ezingu-43. Ukuqeqeshwa kwangemuva kwasebenzisa i-SFT, i-RL yezindawo eziningi enemivuzo eqinisekisekayo, kanye ne-Multi-Teacher On-Policy Distillation kubothisha abangaphezu kweshumi abangochwepheshe. Umgomo womklamo oshiwo yiwo ama-ejenti asebenza isikhathi eside: amaseshini abamba umongo omkhulu, abiza amathuluzi amashumi ezikhathi, futhi asebenza isikhathi eside ngokwanele ukuze umnotho wethokheni ngalinye ubuse ibhili.

02Uzalo lwe-Nemotron, nokuthi kungani umlando ubalulekile

I-Nemotron akuyona into eyenzeke kanye. Ingumphumela obonakalayo wohlelo lwezocwaningo lwe-NVIDIA olusebenza ngakuleyo ndlela kusukela ngo-2024, futhi isizukulwane ngasinye sishiye into imodeli yamanje esaphethe. I-Nemotron-4 340B (Nhlangulana 2024) yathunyelwa nomshini ovulekile wokukhiqiza idatha yokwenziwa, imodeli yayo yomvuzo nesikhiqizi sayo, kanye nokudalula okungajwayelekile: iningi elikhulu ledatha yayo yokuqondaniswa lakhiqizwa yimodeli. Imodeli njengefekthri yedatha kunokuba yisangoma yikho okwenza ukukhishwa kwe-Nemotron kwakamuva kuqhubeke kuthumela amaqoqo edatha kanye nezisindo.

I-Llama-Nemotron — ukucabanga ongakuvala

Umndeni we-Llama-Nemotron ka-2025 (LN-Nano 8B, LN-Super 49B, LN-Ultra 253B) waletha imibono emibili eguge kahle. Owokuqala ubungowokwakhiwa: ukusesha kokwakhiwa kwe-neural okuthi Puzzle kwe-NVIDIA kwakha umtapo wamabhlogo e-transformer acutshungulwe ngamabhlogo, kwabe sekuwahlanganisa ngezinhlelo ze-mixed-integer, kwakhiqiza amamodeli lapho amanye amabhlogo eqa ukunaka ngokuphelele noma ekushintsha ngengqimba elilodwa eliqondile, kanti i-FFN Fusion yagoqa amabhlogo alandelanayo e-feed-forward abe mancane, abanzi futhi ahambisanayo — inciphisa ukujula okulandelanayo kunenani lamapharamitha. I-LN-Ultra yalingana nendawo eyodwa engu-8×H100 kanti, ngokwephepha le-NVIDIA, yadlula i-DeepSeek-R1 ku-GPQA-Diamond (76.0% ngokumelene no-71.5%).

Owesibili wawungowokuziphatha: i-Llama-Nemotron yaba ngumndeni ovulekile wokuqala onenkinobho yokucabanga eguqukayo, imaka idatha yokuqeqesha ngo-detailed thinking on noma u-detailed thinking off emyalelweni wesistimu, ngemivuzo yefomethi ephoqa amathegi e-<think> emodini evuliwe. Isethi eyodwa yezisindo, ukuziphatha okubili, okukhethwa ngesikhathi sokuqagela. Ukulawulwa kwesabelo sokucabanga kwe-Nemotron 3 — nepharamitha ye-reasoning_effort ku-API yethu — kwehla ngqo kulowo msebenzi.

I-Nemotron 3 — ukuphenduka kokuxuba

I-Nemotron 3 yathatha indawo yokususelwa ku-Llama ngokwakhiwa okuxubile okusuka phansi futhi yengeza i-LatentMoE (ukucindezela amathokheni ngaphambi kokuqondisa, i-NVIDIA ethi kuvumela "4x as many experts for the same inference cost"), i-multi-token prediction, kanye nokuqeqeshwa kwangaphambili kwemvelo kwe-NVFP4. I-NVIDIA yashicilela amasampula e-SFT ayizigidi ezingu-50 esewonke, imisebenzi ye-RL eyizigidi ezingu-2 nezindawo ze-RL ezingu-55 kanye nezisindo — akudingeki ukuthi wethembe izwi lomthengisi ngalokho imodeli eyakufundiswa, ngoba impahla yokufundisa iyalandeka.

03Yini oyithengayo ngokwakhiwa okuxubile

I-KV cache ye-transformer ikhula ngokulingana nobude bochungechunge, yingakho ukuqagela komongo omude kubiza kakhulu enkumbulweni kunasekubaleni. Amangqimba e-Mamba-2 angamamodeli e-state-space aphethe isimo esiphindaphindayo esinosayizi ongaguquki, ngakho izindleko zenkumbulo zawo kuthokheni ngalinye azikhuli nomlando; lokho angakwazi kahle ukukhumbula okunembayo okuhlobene. Okuxubile kugcina i-Mamba-2 kwingxenye enkulu yokucutshungulwa kochungechunge kanye nengxenye encane yamangqimba angempela okunaka ukuze kubuyiswe lokho amangqimba e-state-space awaphutha kukho, ngesilinganiso esiqinile sezinhloko zombuzo ezingu-64 nezinhloko ze-KV ezimbili esinciphisa ikheshi nakakhulu. I-NVIDIA ibika u-94.7% ku-RULER emathokhenini ayisigidi — ukuhlolwa kokuthola inaliti okwakhelwe ukubamba amamodeli athi anewindi angakwazi ukulisebenzisa.

Umphumela ukukhishwa: i-NVIDIA ithi ukukhishwa kokuqagela kufika cishe ku-6× kunamamodeli afanayo ekunembeni okulinganayo, kanti i-Artificial Analysis yakala ngokuzimela amathokheni angu-140.3 ngomzuzwana nesikhathi sethokheni lokuqala esingu-1.33 yomzuzwana — indawo yesi-7 kumamodeli angu-89. Ekusebenzeni kwe-ejenti kwemizuliswano engu-200 lowo ngumehluko phakathi kokuqeda ntambama nokungaqedi nhlobo.

04Kusho ukuthini u-"3BIT-REAP", ngeqiniso

Isijobelelo asisona umhlobiso — sikutshela ncamashi ukuthi kwenziwani esisekelweni ukuze silingane.

I-REAP — Router-weighted Expert Activation Pruning — ivela ku-Cerebras Research futhi yamukelwa ku-ICLR 2026. Ku-MoE esabalele i-router ithumela ithokheni ngalinye kochwepheshe abambalwa; emsebenzini wangempela abanye bashaywa njalo kanti abanye cishe abalokothi. I-REAP inikeza uchwepheshe ngamunye amaphuzu ngesisekelo sokubaluleka esihlanganisa ukuthi i-router imvula ngamandla amangakanani kanye nokuthi mikhulu kangakanani iminikelo yokukhishwayo kwakhe, bese ikhipha ngokuphelele abehluleka ngokuphelele — i-Cerebras ibika ukuthena kufika ku-50% wochwepheshe kumamodeli amakhulu njengalawo anamapharamitha ayisigidintathu kuyilapho igcina ikhwalithi eyisisekelo, futhi ithola ukuthi emisebenzini yokukhiqiza njengekhodi, ukuthena kudlula ukuhlanganisa. Ukukalwa kwamabhithi amathathu bese kucindezela lokho okusalayo. Thena kuqala, ukale kwesibili, ukuze isabelo samabhithi siye kochwepheshe abaphethe ithrafikhi esikhundleni sokusakazwa phezu kwabafile.

Ukucindezelwa akumahhala, futhi ngeke senze sengathi kunjalo: ukukalwa okuqinile kuqala kuvele njengokwehla okuncane ekulandeleni ifomethi eqondile nasezibalweni zeketango elide, hhayi njengokungezwakali okubonakalayo. Lapho izinga elithathu ehlala khona:

Isijobelelo	Siyini	Lapho sisisebenzisa khona
3BIT-REAP	Izisindo zamabhithi amathathu + ukuthenwa kochwepheshe	Izisekelo ezinkulu kunazo zonke, lapho okunye konke kungalingani ngesivinini
W4A16-AUTOROUND-REAP	Izisindo zamabhithi amane / ukusebenza kwamabhithi ayi-16, ukugoba okufundiwe kwe-AutoRound, + ukuthena	Izisekelo ezimaphakathi; i-AutoRound ibuyisa ukunemba kunokugoba okulula
W8A16	Izisindo zamabhithi ayisishiyagalombili / ukusebenza kwamabhithi ayi-16	Ukwethembeka okuphezulu kunakho konke; izisekelo ezincane ngokwanele ukuze zingadingi ukucindezelwa okuqinile

I-Nemotron 3 Ultra engu-550B ingqo esigabeni sokuqala. Uma umsebenzi wakho uzwela ngokungajwayelekile kokukhishwayo okunezinombolo eziqondile emaketangweni amade, inkomba yocwaningo ibala izinga le-W8A16 — u-MiMo-V2.5-Pro-W8A16 no-Hy3-W8A16 — kuphansi ka-$0.50/$2.00. Ukushintshana kwangempela, hhayi umgudu wokuthengisa: imodeli encane, ukucindezelwa okuncane.

05Kusho ukuthini u-"ongavinjelwe" kulesi siphetho

Ngokuqondile: alukho ungqimba lokwenqaba futhi akukho ukuhlungwa kokuqukethwe okusetshenziswa kokukhishwayo. Asikho isihlukanisi esihlola umyalelo wakho bese sibuyisela ukwenqaba okulungiselelwe ngaphambili, futhi impendulo ayihlolwa ngaphambi kokuba ifike emsingeni wakho. Okutholayo ukuziphatha kwemodeli uqobo.

Yilapha lapho uzalo oluvulekile luyeka khona ukuba yinto engabalulekile. Ngenxa yokuthi i-NVIDIA yashicilela idatha yokuqeqeshwa kwangemuva nezindawo ze-RL, ukuthambekela kwemodeli kungahlolwa ngokomqondo kunokuba kube yinganekwane. Futhi ngenxa yokuthi i-OpenMDW-1.1 yilayisensi evumelanayo ye-Linux Foundation ehlanganisa ukwakhiwa, amapharamitha, imibhalo nesofthiwe, umqhubi ozimele angakhonza ukukhishwa okuguquliwe, okungahlungiwe kuma-hardware akhe ngaphandle kokuthi ilayisensi ibe yisithiyo. Amalayisensi amaningi amamodeli "avulekile" awakuvumeli lokho. Leli liyakuvumela. Umehluko uyavela lapho ukwenqaba okuzenzakalelayo kuyibhugi, hhayi isici sokuphepha:

• Ucwaningo lokuphepha — ukuhlaziywa kobuthakathaka, ukucabanga nge-exploit, ukuziphatha kwe-malware, amathuluzi eqembu elibomvu, ngaphandle kokuphikisana ngokuthi umbuzo uyavunyelwa yini.
• Ikhemistri, ibhayoloji nefamakholoji — izibalo zomthamo, izindlela zokusabela, ubuthi, izindlela zokwakha izinto lesiphetho esihlungiwe esizenqabela ngokuthinta igama nje.
• I-ML ephikisanayo — okubi okulukhuni, amaqoqo e-jailbreak nezinhlelo zokuhlukaniswa kokuhlukumeza kwezihlukanisi ezidinga ukuthi zibone into yangempela.
• Umbhalo womtholampilo, wezomthetho nowe-forensic — impahla yamacala ecacile ngokwemvelo bese yenqatshwa ngenxa yaleso sizathu ncamashi.

Okungavinjelwe akusho ukungabi nesibopho. Ungumnikazi walokho okukhiqizayo, futhi Inqubomgomo Yethu Yokusebenzisa Ngokuzibophezela kanye Nemigomo Yesevisi kubusa i-akhawunti. Ukususa ungqimba lokwenqaba kususa ukuqagelwa ngenhloso; akususi izibopho zakho ngaphansi komthetho osebenza kuwe.

06I-STEM nokucabanga: ashoni amanani ashicilelwe

Ukugxila kokuqeqeshwa kwe-Nemotron kube yizibalo, ikhodi, isayensi nokulandela imiyalelo kusukela iqoqo ledatha lokuqeqeshwa kwangemuva le-Llama-Nemotron lakhelwa ngqo phezu kwalezo zici ezine. Lokho i-NVIDIA ekubikayo ngesisekelo se-Nemotron 3 Ultra — amanani abikwe ngumthengisi ngemodeli engacindezelwe, esingasazange siwaqhube:

Izixwayiso ezimbili ezibaluleke ukwedlula amanani. Inani le-SWE-Bench Verified liyibanga — i-NVIDIA ibika u-65.0% kuya ku-70.4% kuya ngokuthi yimuphi umshini we-ejenti oyiqhubayo (Pi, OpenHands, Hermes, OpenCode, Mini SWE Agent), futhi ukusabalala kwamaphuzu amahlanu ngenxa yesakhiwo sokweseka kuphela kukhombisa ukuthi amaphuzu e-ejenti axhomeke kangakanani kumshini kunakwimodeli. Futhi i-Artificial Analysis ibeke i-Nemotron 3 Ultra ku-48 ku-Intelligence Index yayo, indawo yesi-9 kumamodeli angu-89 — inamandla, kodwa hhayi ephezulu kunazo zonke, futhi ukukala kwabo okuzimele yikho esingakubeka phezu kwanoma yisiphi isilayidi somthengisi.

Esizukulwaneni sangaphambili, iphepha le-Llama-Nemotron le-NVIDIA libika i-LN-Ultra ku-76.0% ku-GPQA-Diamond imelene no-71.5% we-DeepSeek-R1 no-43.4% we-Llama-3.1-405B-Instruct, kanye no-80.8% ku-AIME 2024 — imodeli ehlukile, efakiwe ngoba isayensi yezinga leziqu nezibalo zomncintiswano kube yikho okuhloswe ngumndeni izizukulwane ezimbili.

Amanani okuhlolwa njengoba eshicilelwe yi-NVIDIA (Nhlangulana 2026) nge-Nemotron 3 Ultra engacindezelwe, kanye nayi-Artificial Analysis ekukalweni okuzimele. Okwakhiwe okucindeziwe kungahluka. Asishicileli amanani okuhlolwa esingazange siwaqinisekise noma esingakwazi ukuwanikeza umthombo.

07Intengo, nesixwayiso sokukhuluma kakhulu

Ku-$0.75 ngethokheni eyisigidi engenayo no-$3.30 ngethokheni eyisigidi ephumayo, i-Nemotron3Ultra ihlala kwingxenye yesithathu ephansi yekhathalogu yethu kokungenayo naphakathi kokuphumayo. Izibonelo ezisebenzayo, ngoba intengo ngesigidi kunzima ukuyizwa:

Isimo sokubiza	Okungenayo	Okuphumayo	Izindleko
Umzuliswano ojwayelekile we-RAG	50,000	4,000	$0.051
Ukuhlaziya idokhumenti ende	150,000	8,000	$0.139
Ukubiza kosayizi ophezulu	262,144	16,384	$0.251

Isicelo esinomongo ogcwele nokukhishwayo okugcwele — lonke ithokheni isiphetho esizolamukela nalo lonke elizolikhiqiza — sibiza cishe amasenti angu-25. Isimo esifanayo ku-Kimi-K3-3BIT-REAP, imodeli yethu ebiza kakhulu ku-$3.83/$19.12, sibiza cishe u-$1.32, okungu-5.3× ngaphezulu.

Manje isixwayiso, futhi singesangempela. I-Artificial Analysis yakala i-Nemotron 3 Ultra ikhiqiza amathokheni okukhishwayo angu-2.3× ngaphezu kwesilinganiso samamodeli ayintanga kuwo wonke amasethi awo okuhlola. Amamodeli okucabanga akukhokhisa ngokucabanga, futhi le ecabanga isikhathi eside; inzuzo yentengo ngethokheni iyanyamalala ibe yilahleko ngomsebenzi uma imodeli icabanga isikhathi esiphindwe kathathu ukuze ifinyelele empendulweni efanayo. Ikhambi lakhelwe ngaphakathi futhi livezwe ku-API yethu: i-Nemotron 3 yaqeqeshwa ngemuva ukulawula isabelo sokucabanga, futhi u-reasoning_effort uyasekelwa lapha — i-NVIDIA ichaza imodi yomzamo ophansi esebenzisa amathokheni okucabanga amancane kakhulu ngokulahlekelwa ukunemba okusenanini elilodwa. Ekuhlukaniseni ngezigaba, ekukhipheni ulwazi nasekuqondiseni, wehlise futhi umnotho uyashintsha ngendlela ebonakalayo; ebufakazini noma ekuguqulweni kwamafayela amaningi, ukhuphule bese ukhokha ngenhloso. Kala amathokheni akho ngomsebenzi kunokufunda uhlu lwamanani bese ucabangela.

08Ukuyibiza: izilimi ezintathu, imodeli eyodwa

Yonke imodeli esangweni iphendula ezindaweni ezintathu ze-API ezisakazayo, ngakho ungakhomba iklayenti elikhona kuyo ngaphandle kokubhala kabusha ukuhlanganiswa kwakho: /v1/chat/completions (ulimi lwe-OpenAI), /v1/messages (ulimi lwe-Anthropic), kanye no-/v1/responses.

curl https://api.shannon-ai.com/v1/chat/completions \
  -H "Authorization: Bearer $SHANNON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Nemotron3Ultra-3BIT-REAP",
    "reasoning_effort": "medium",
    "stream": true,
    "messages": [
      {"role": "user", "content": "Derive the worst-case cache behavior of this loop nest."}
    ]
  }'

Isifinyezo samakhono, ukuze kungabikho okukumangazayo ngesikhathi sokuhlanganisa:

Ikhono	Nemotron3Ultra · 3BIT-REAP
Iwindi lomongo	amathokheni angu-262,144
Okukhishwayo okuphezulu / isicelo	amathokheni angu-16,384
Ukusakaza	Yebo
Ukubiza amathuluzi	Yebo
Imikhondo yokucabanga	Yebo
reasoning_effort	Yebo
json_schema (okuqinile)	Yebo
response_format	Yebo
Umbono / okufakwayo okuyisithombe	Cha — umbhalo kuphela
Ungqimba lokwenqaba	Alukho

Umugqa wombono ubamba abantu: i-Nemotron3Ultra ingeyombhalo kuphela. Uma umshini wakho uthumela izithombe, sebenzisa u-MiniMax-M3-3BIT-REAP noma u-MiMo-V2.5-W8A16 ku-$0.50/$2.00, noma u-Kimi-K2.6-W4A16-AUTOROUND-REAP ku-$0.78/$3.67 — zonke zikwazi umbono, zonke zingavinjelwe, zonke zisemongweni ofanayo wamathokheni angu-262,144. Ireferensi ephelele yamapharamitha iku-mibhalo ye-API.

09Ihlalaphi i-Nemotron3Ultra phakathi kwayishumi nambili

Ikhathalogu ingamamodeli ayishumi nambili angavinjelwe abelana ngewindi lomongo elifanayo, izilimi ezintathu ezifanayo nokungabikho okufanayo kongqimba lokwenqaba; ahluka ngesisekelo, ngentengo nangamakhono. Indawo ye-Nemotron icacile: inkabi yokucabanga yozalo oluvulekile — hhayi eshibhe kunazo zonke, hhayi enamandla kunazo zonke, kodwa yiyo enomkhondo wamaphepha ovulekile omningi ngemuva kokuziphatha kwayo, elungiselelwe kakhulu i-STEM nama-ejenti.

Imodeli	Engenayo / 1M	Ephumayo / 1M	Umbono	Yikhethe uma
MiniMax-M3 · 3BIT-REAP	$0.50	$2.00	Yebo	Inketho eshibhe kunazo zonke enombono
MiMo-V2.5-Pro · W8A16	$0.50	$2.00	Cha	Izinga lokukala elethembeke kunazo zonke
GLM-5.2 · 3BIT-REAP	$0.73	$2.34	Cha	Ukucabanga kwesisekelo esikhulu okushibhe kunakho konke
Nemotron3Ultra · 3BIT-REAP	$0.75	$3.30	Cha	I-STEM, ama-ejenti, uzalo oluvulekile
Kimi-K2.6 · W4A16-AUTOROUND-REAP	$0.78	$3.67	Yebo	Umbono + ukwethembeka kwe-AutoRound
DeepSeek-V4-Pro · 3BIT-REAP	$1.95	$3.90	Cha	Ukuhlaziya okujulile kokushaya kanye
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	Yebo	Uphahla lwamandla oluphezulu ngokuphelele

Kuboniswe eziyisikhombisa kweziyishumi nambili. Ikhathalogu ephelele — kufaka i-Laguna-S-2.1, i-inkling, i-DeepSeek-V4-Flash, i-MiMo-V2.5 kanye ne-Hy3 — iku-mibhalo ye-API. Qaphela ukuthi i-Laguna-S-2.1 ayikusekeli ukukhishwayo okuhlelekile futhi i-inkling ayimsekeli u-json_schema; i-Nemotron3Ultra isekela kokubili.

Umakhelwane oseduze ngentengo ngu-GLM-5.2-3BIT-REAP ku-$0.73/$2.34, futhi ukuqhathanisa okuqotho yilokhu: i-GLM-5.2 ishibhile ngethokheni, ikakhulukazi kokuphumayo. I-Nemotron iletha iqoqo ledatha lokuqeqesha elishicilelwe, isishibilikisi esisobala sesabelo sokucabanga, kanye nomlando womndeni oqondiswe kwisayensi yezinga leziqu. Ukucabanga okuvamile kwevolumu ephezulu: sebenzisa eshibhile. Okugxile ku-STEM, okwe-ejenti, noma okungaphansi kokuhlolwa okubuza ukuthi ukuziphatha kwemodeli kwavelaphi: i-Nemotron iyawuzuza umehluko.

10Imibuzo ebuzwa njalo

---

Imithombo: Umbiko wobuchwepheshe we-NVIDIA Nemotron 3 Ultra · Ibhulogi yonjiniyela ye-NVIDIA (Nhlangulana 2026) · Amanothi okwakhiwa kwe-Nemotron 3 Super · Llama-Nemotron: Efficient Reasoning Models · Cerebras Research, REAP. Ukuhlolwa kwesisekelo kubikwe yi-NVIDIA ngemodeli engacindezelwe ngaphandle uma kunikezwe i-Artificial Analysis; okwakhiwe okucindeziwe kungahluka. Amanani nezakhi zamanje kusukela ngoMandulo 5, 2026.