API na Nemotron Marar Tantancewa

Yawancin samfura "buɗaɗɗu" nauyi ne kawai aka buɗe, ba komai kuma — ɗamarar sigogi, lasisi, da shiru game da inda lambobin suka fito. Nemotron ya kasance banda na dindindin tsawon shekaru biyu: NVIDIA yana fitar da nauyin, bayanan horo na baya, muhallan RL da girke-girken. Wannan yana da muhimmanci ga ƙarshen hanya marar tantancewa ta hanyar da ake da sauƙin kuskurewa. Samfurin da aka wallafa bayanan daidaitawarsa shi ne wanda za ka iya yin tunani a kan ɗabi'ar ƙinsa maimakon ka yi tsammani — kuma wanda wani ɓangare na uku zai iya tura shi bisa doka ba tare da matatar abun ciki a maƙogwaronsa ba.

01Menene Nemotron 3 Ultra?

Nemotron 3 Ultra shi ne mafi girma a cikin dangin Nemotron 3 na NVIDIA, an fitar da shi zuwa Hugging Face a 4 ga Yuni, 2026. Rahoton fasaha na NVIDIA ya kwatanta shi da cewa "samfurin harshe na Mixture-of-Experts Hybrid Mamba-Attention mai jimlar sigogi biliyan 550 da biliyan 55 masu kunnuwa" — kusan kashi 10% na hanyar sadarwar ke aiki a kan kowane token, wanda shi ne ya sa samfuri mai rabin tiriliyan sigogi ya zama mai yiwuwar hidima kwata-kwata.

Tsarin da aka bayar da rahotonsa shi ne shimfidu 108 a girman samfuri 8,192, yana haɗa shimfidun sararin yanayi na Mamba-2 da ƙaramin adadin cikakkun shimfidun kulawa, masana 512 a kowace shimfidar MoE da turawa ta top-22. Horo na farko ya gudana a kan kusan tokens tiriliyan 20 na rubutu a matakai biyu — na farko mai karkata ga bambanci, sannan mai karkata ga inganci — da tsayawa a Satumba 2025, harsuna 12 na ɗabi'a da harsunan shirye-shirye 43. Horo na baya ya yi amfani da SFT, RL a muhalli da dama tare da sakamako masu tabbatuwa, da Multi-Teacher On-Policy Distillation a kan malamai ƙwararru fiye da goma. Manufar ƙirƙirar da aka bayyana ita ce wakilai masu dogon gudu: zaman da suke riƙe babban mahalli, suna kiran kayan aiki sau da dama, kuma suna gudana har tsawon lokacin da tattalin arziki na kowane token ya mamaye lissafin kuɗi.

02Zuriyar Nemotron, da dalilin da ya sa asalin yake da muhimmanci

Nemotron ba abu ne na sau ɗaya ba. Fitowa ce bayyananniya ta shirin bincike na NVIDIA da yake tafiya a wannan hanyar tun 2024, kuma kowane tsara ya bar wani abu da samfurin yanzu yake ci gaba da ɗauka. Nemotron-4 340B (Yuni 2024) ya fito da buɗaɗɗiyar hanyar samar da bayanai na roba, samfurin sakamakonsa da mai samar da shi, tare da wani bayani mai ban mamaki: mafi yawan bayanan daidaitawarsa samfuri ne ya samar da su. Samfuri a matsayin masana'antar bayanai maimakon mai duba shi ne dalilin da ya sa sakin Nemotron na baya suke ci gaba da fitar da bayanai tare da nauyi.

Llama-Nemotron — tunanin da za ka iya kashewa

Dangin Llama-Nemotron na 2025 (LN-Nano 8B, LN-Super 49B, LN-Ultra 253B) ya bayar da ra'ayoyi biyu da suka daɗe suna aiki. Na farko na gine-gine ne: binciken gine-ginen hanyar sadarwa na Puzzle na NVIDIA ya gina ɗakin karatu na toshen transformer da aka tace toshe-toshe sannan ya haɗa su da shirye-shirye na mixed-integer, yana samar da samfura inda wasu toshe suke tsallake kulawa gaba ɗaya ko su maye gurbinta da shimfida ɗaya kai tsaye, yayin da FFN Fusion ta durƙusar da toshen feed-forward masu biye da juna zuwa ƙalilan, masu faɗi, masu aiki tare — tana rage zurfin jere maimakon adadin sigogi. LN-Ultra ya shiga cikin kumburi guda na 8×H100 alhali, bisa takardar NVIDIA, ya zarce DeepSeek-R1 a GPQA-Diamond (76.0% da 71.5%).

Na biyu na ɗabi'a ne: Llama-Nemotron shi ne dangi buɗaɗɗe na farko mai maɓallin tunani mai motsi, yana alamta bayanan horo da detailed thinking on ko detailed thinking off a cikin umarnin tsarin, tare da sakamakon tsari da suke tilasta alamun <think> a yanayin kunne. Nauyi guda, ɗabi'u biyu, ana zaɓarsu a lokacin tunani. Sarrafa kasafin tunani na Nemotron 3 — da sigar reasoning_effort a kan API ɗinmu — sun fito kai tsaye daga wannan aikin.

Nemotron 3 — juyi zuwa haɗe

Nemotron 3 ya maye gurbin zuriyar Llama da gine-gine haɗe daga tushe kuma ya ƙara LatentMoE (matsa tokens kafin turawa, wanda NVIDIA ya ce yana ba da damar "masana sau 4 a kan kuɗin tunani guda"), hasashen tokens da yawa, da horo na farko na asali a NVFP4. NVIDIA ya wallafa jimlar samfuran SFT miliyan 50, ayyukan RL miliyan 2 da muhallan RL 55 tare da nauyin — ba dole ne ka amince da maganar mai kera game da abin da aka koya wa samfurin ba, saboda ana iya sauke kayan koyarwar.

03Abin da gine-ginen haɗe yake saya maka

Ma'ajiyar KV ta transformer tana girma kai tsaye da tsawon jerin, wanda shi ne ya sa tunani a dogon mahalli yake tsada a ƙwaƙwalwa kafin ya yi tsada a lissafi. Shimfidun Mamba-2 samfuran sararin yanayi ne masu ɗauke da yanayi mai maimaituwa mai girma tsayayye, don haka kuɗin ƙwaƙwalwarsu na kowane token ba ya girma da tarihi; abin da ba sa iya yi shi ne dawo da bayani na haɗi daidai. Haɗin yana riƙe Mamba-2 don yawancin sarrafa jerin tare da ƙaramin adadin shimfidun kulawa na gaskiya don dawo da bayanin da shimfidun sararin yanayi suke gazawa a kai, da rabon kai 64 na tambaya zuwa kai 2 na KV mai tsauri wanda yake ƙara rage ma'ajiyar. NVIDIA ya bayar da rahoton 94.7% a RULER a tokens 1M — ma'aunin dawo da allura da aka gina don kama samfuran da suke iƙirarin tagar da ba za su iya amfani da ita ba.

Sakamakon shi ne yawan aiki: NVIDIA yana iƙirarin har zuwa kusan sau 6 mafi girma na yawan tunani fiye da samfura makamantansa a daidaito iri ɗaya, kuma Artificial Analysis ya auna da kansa tokens 140.3 a kowane daƙiƙa da daƙiƙa 1.33 na lokaci zuwa token na farko — na 7 daga cikin samfura 89. A gudun wakili na zagaye 200 wannan shi ne bambanci tsakanin kammalawa da yamma da rashin kammalawa kwata-kwata.

04Menene ma'anar "3BIT-REAP" a gaskiya

Ƙarin ba ado ba ne — yana gaya maka daidai abin da aka yi wa tushen don ya shiga.

REAP — Router-weighted Expert Activation Pruning — ya fito daga Cerebras Research kuma an karɓe shi zuwa ICLR 2026. A cikin MoE mai tsattsauri router yana aika kowane token zuwa ƴan masana; a kan ainihin aiki wasu ana kai gare su koyaushe wasu kuma kusan ba a taɓa kai gare su ba. REAP yana ba kowane masani maki bisa ma'aunin muhimmanci da yake haɗa ƙarfin da router yake kai gare shi da girman gudummawar fitarwarsa, sannan ya cire waɗanda suka fi ƙaranci kai tsaye — Cerebras ya bayar da rahoton datsa har zuwa kashi 50% na masana daga samfura masu girman sigogi tiriliyan 1 alhali ya kiyaye ingancin asali gaba ɗaya, kuma ya gano cewa ga ayyukan samarwa kamar lamba, datsawa ta fi haɗawa. Sannan matsi na 3-bit yana matsa abin da ya tsira. Datse da farko, matsa na biyu, don kasafin bit ya tafi ga masanan da suke ɗauke da zirga-zirga maimakon a shafa shi a kan waɗanda suka mutu.

Matsawa ba kyauta ba ce, kuma ba za mu yi kamar ba haka ba: matsi mai tsauri yana bayyana da farko a matsayin ƙananan raguwa a bin tsari daidai da lissafi mai dogon sarƙa, ba a matsayin rashin fahimta bayyananne ba. Ga inda matakan uku suke:

Ƙari	Menene shi	Inda muke amfani da shi
3BIT-REAP	Nauyi na 3-bit + datsa masana	Manyan tushe, inda babu wani abu da zai shiga a kan sauri
W4A16-AUTOROUND-REAP	Nauyi na 4-bit / kunnuwa na 16-bit, zagayen da AutoRound ya koya, + datsawa	Tushe masu matsakaicin girma; AutoRound yana dawo da daidaito a kan zagaye zuwa mafi kusa mai sauƙi
W8A16	Nauyi na 8-bit / kunnuwa na 16-bit	Mafi girman aminci; tushe ƙanana har ba sa buƙatar matsi mai tsauri

Nemotron 3 Ultra a 550B yana daidai a rukunin farko. Idan aikinka yana da matuƙar hankali ga fitarwar lamba daidai a kan dogayen sarƙoƙi, fihirisar bincike tana jera matakin W8A16 — MiMo-V2.5-Pro-W8A16 da Hy3-W8A16 — a ƙasan $0.50/$2.00. Ciniki na gaskiya, ba wata magudi ba: ƙaramin samfuri, ƙarancin matsawa.

05Menene ma'anar "marar tantancewa" a wannan ƙarshen hanyar

A takaice: babu shimfidar ƙi kuma babu matatar abun ciki da aka sa a kan fitarwa. Babu wani mai rarrabewa da yake duba umarninka sannan ya mayar da ƙi da aka riga aka rubuta, kuma ba a duba amsar kafin ta isa rafinka ba. Abin da ka samu shi ne ɗabi'ar samfurin da kansa.

A nan ne zuriyar buɗaɗɗiya ta daina zama bayani mara amfani. Saboda NVIDIA ya wallafa bayanan horo na baya da muhallan RL, ana iya duba halayen samfurin bisa ƙa'ida maimakon a dogara da jita-jita. Kuma saboda OpenMDW-1.1 lasisi ne mai saukin kai na Linux Foundation da ya ƙunshi gine-gine, sigogi, takardu da software, mai gudanarwa mai zaman kansa zai iya yin hidimar tura da aka canja, mara matata, a kan na'urorinsa ba tare da lasisin ya zama cikas ba. Lasisin samfura "buɗaɗɗu" da dama ba sa ba da izinin hakan. Wannan yana bayarwa. Bambancin yana bayyana inda ƙi ta atomatik matsala ce, ba fasalin tsaro ba:

• Binciken tsaro — nazarin rauni, tunani a kan hanyar amfani da rauni, ɗabi'ar malware, kayan aikin red-team, ba tare da jayayya a kan ko an halatta tambayar ba.
• Sinadarai, ilimin halittu da magunguna — lissafin adadin magani, hanyoyin sarrafawa, ilimin guba, hanyoyin haɗawa da ƙarshen hanya mai matata yake ƙi a kan taɓa kalma kawai.
• ML na gaba da gaba — misalai marasa kyau masu wuya, tarin jailbreak da rabe-raben cin zarafi ga masu rarrabewa da ke buƙatar sun ga ainihin abin.
• Rubutun asibiti, shari'a da bincike na shaida — kayan shari'a da suke da tsanani ta ɗabi'arsu kuma ake ƙinsu saboda daidai wannan dalilin.

Rashin tantancewa ba ya nufin rashin alhaki. Kai ne ke da abin da ka samar, kuma Manufarmu ta Amfani da Alhaki da Sharuɗɗan Hidima suna gudanar da asusun. Cire shimfidar ƙi yana cire tsammani game da niyya; ba ya cire wajibcinka a ƙarƙashin dokar da ta shafe ka ba.

06STEM da tunani: abin da lambobin da aka wallafa suke faɗa

Abin da horon Nemotron ya mayar da hankali a kai shi ne lissafi, lamba, kimiyya da bin umarni tun lokacin da aka gina bayanan horo na baya na Llama-Nemotron a kan waɗannan gatura huɗu daidai. Abin da NVIDIA ya bayar da rahoto ga tushen Nemotron 3 Ultra — lambobin da mai kera ya bayar da rahotonsu a kan samfurin da ba a matsa ba, waɗanda ba mu gudanar da su ba:

Gargaɗi biyu sun fi lambobin darajar. Lambar SWE-Bench Verified iyaka ce — NVIDIA ya bayar da rahoton 65.0% zuwa 70.4% dangane da harsashin wakilin da ya gudanar da shi (Pi, OpenHands, Hermes, OpenCode, Mini SWE Agent), kuma bambanci na maki biyar daga tsarin gini kaɗai yana nuna yadda makin wakilai suke dogara ga harsashin maimakon samfurin. Kuma Artificial Analysis ya sanya Nemotron 3 Ultra a 48 a Ma'aunin Hankalinsa, na 9 daga cikin samfura 89 — mai ƙarfi, ba a saman fagen ba, kuma ma'aunin da suka yi da kansu shi ne wanda za mu fi ba shi nauyi a kan kowace nunin faifai ta mai kera.

A tsara ta baya, takardar Llama-Nemotron ta NVIDIA ta bayar da rahoton LN-Ultra a 76.0% a GPQA-Diamond a kan 71.5% na DeepSeek-R1 da 43.4% na Llama-3.1-405B-Instruct, tare da 80.8% a AIME 2024 — samfuri daban, an haɗa shi saboda kimiyyar matakin digiri na biyu da lissafin gasa sun kasance abin da dangin ke nufi tsawon tsararraki biyu.

Lambobin ma'auni kamar yadda NVIDIA ya wallafa su (Yuni 2026) a kan Nemotron 3 Ultra da ba a matsa ba, da kuma Artificial Analysis ga ma'aunan da aka yi da kai. Ginin da aka matsa na iya bambanta. Ba ma wallafa lambobin ma'auni da ba mu tabbatar da su ba ko ba za mu iya danganta su ba.

07Farashi, da gargaɗin yawan magana

A $0.75 kowane tokens miliyan ɗaya na shigarwa da $3.30 kowane tokens miliyan ɗaya na fitarwa, Nemotron3Ultra yana zaune a ƙasan sulusin katalogin mu a shigarwa da tsakiya a fitarwa. Misalan aiki, saboda farashin kowane miliyan yana da wuyar ji:

Siffar kira	Shigarwa	Fitarwa	Kuɗi
Zagayen RAG na yau da kullum	50,000	4,000	$0.051
Nazarin dogon takarda	150,000	8,000	$0.139
Kira mai matsakaicin girma	262,144	16,384	$0.251

Buƙata mai cikakken mahalli da cikakkiyar fitarwa — kowane token da ƙarshen hanyar zai karɓa da kowane token da zai samar — tana kashe kusan cent 25. Siffar guda a kan Kimi-K3-3BIT-REAP, samfurin mu mafi tsada a $3.83/$19.12, tana kashe kusan $1.32, ko sau 5.3 fiye da haka.

Yanzu ga gargaɗin, kuma na gaskiya ne. Artificial Analysis ya auna Nemotron 3 Ultra yana samar da tokens na fitarwa sau 2.3 fiye da matsakaicin takwarorinsa a kan jerin ma'aunansu. Samfuran tunani suna caje ka don tunanin, kuma wannan yana yin tunani mai tsawo; ribar farashi a kowane token tana ɓacewa ta zama asara a kowane aiki idan samfurin ya yi tunani sau uku fiye da haka don ya kai amsa ɗaya. Maganin an gina shi ciki kuma an buɗe shi a API ɗinmu: an horar da Nemotron 3 a baya don sarrafa kasafin tunani, kuma ana goyon bayan reasoning_effort a nan — NVIDIA ya kwatanta yanayin ƙoƙari mai ƙasa da yake amfani da ƙarancin tokens na tunani sosai a kan ciniki na daidaito na maki ɗaya. A rarrabewa, cirar bayani da turawa, ka rage shi kuma tattalin arziki zai canja sosai; a kan hujja ko sake gina fayiloli da yawa, ka ɗaga shi kuma ka biya kuɗinsa da gangan. Ka auna tokens ɗinka na kowane aiki maimakon ka karanta jerin farashi ka yi tsammani.

08Kiransa: yaruka uku, samfuri guda

Kowane samfuri a ƙofar shiga yana amsawa a kan filayen API guda uku masu streaming, don haka kana iya nuna wa abokin ciniki da kake da shi ba tare da sake rubuta haɗin ka ba: /v1/chat/completions (yaren OpenAI), /v1/messages (yaren Anthropic), da /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."}
    ]
  }'

Taƙaitaccen iyawa, don kada wani abu ya ba ka mamaki a lokacin haɗawa:

Iyawa	Nemotron3Ultra · 3BIT-REAP
Taga mahalli	tokens 262,144
Matsakaicin fitarwa / buƙata	tokens 16,384
Streaming	Eh
Kiran kayan aiki	Eh
Alamun tunani	Eh
reasoning_effort	Eh
json_schema (mai tsauri)	Eh
response_format	Eh
Gani / shigar hoto	A'a — rubutu kaɗai
Shimfidar ƙi	Babu

Layin gani yana kama mutane: Nemotron3Ultra rubutu kaɗai ne. Idan hanyar aikinka tana aika hotuna, yi amfani da MiniMax-M3-3BIT-REAP ko MiMo-V2.5-W8A16 a $0.50/$2.00, ko Kimi-K2.6-W4A16-AUTOROUND-REAP a $0.78/$3.67 — duk suna iya gani, duk marasa tantancewa, duk a kan mahalli na tokens 262,144 guda. Cikakken bayanin sigogi yana cikin takardun API.

09Inda Nemotron3Ultra yake a cikin goma sha biyun

Katalogin samfura goma sha biyu ne marasa tantancewa da suke raba taga mahalli guda, yaruka uku guda da rashin shimfidar ƙi guda; sun bambanta a tushe, farashi da iyawa. Matsayin Nemotron takamaimai ne: dokin aikin tunani mai zuriya buɗaɗɗiya — ba mafi rahusa ba, ba mafi iyawa ba, amma wanda ya fi kowa takardu na jama'a a bayan ɗabi'arsa, kuma aka fi kera shi don STEM da wakilai.

Samfuri	Shiga / 1M	Fita / 1M	Gani	Zaɓe shi don
MiniMax-M3 · 3BIT-REAP	$0.50	$2.00	Eh	Zaɓi mafi rahusa mai iya gani
MiMo-V2.5-Pro · W8A16	$0.50	$2.00	A'a	Matakin matsi mafi girman aminci
GLM-5.2 · 3BIT-REAP	$0.73	$2.34	A'a	Tunanin babban tushe mafi rahusa
Nemotron3Ultra · 3BIT-REAP	$0.75	$3.30	A'a	STEM, wakilai, zuriya buɗaɗɗiya
Kimi-K2.6 · W4A16-AUTOROUND-REAP	$0.78	$3.67	Eh	Gani + amincin AutoRound
DeepSeek-V4-Pro · 3BIT-REAP	$1.95	$3.90	A'a	Zurfin nazari na harbi guda
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	Eh	Cikakken rufin iyawa

Bakwai daga cikin goma sha biyu aka nuna. Cikakken katalogin — haɗe da Laguna-S-2.1, inkling, DeepSeek-V4-Flash, MiMo-V2.5 da Hy3 — yana cikin takardun API. Ka lura cewa Laguna-S-2.1 ba ya goyon bayan fitarwa mai tsari kuma inkling ba ya goyon bayan json_schema; Nemotron3Ultra yana goyon bayan duka biyun.

Maƙwabci mafi kusa bisa farashi shi ne GLM-5.2-3BIT-REAP a $0.73/$2.34, kuma kwatanci na gaskiya shi ne: GLM-5.2 ya fi rahusa a kowane token, musamman a fitarwa. Nemotron yana kawo tarin horo da aka wallafa, maɓallin kasafin tunani bayyananne, da tarihin dangi da aka nufa ga kimiyyar matakin digiri na biyu. Tunani na kowane fanni mai girma mai yawa: gudanar da wanda ya fi rahusa. Mai nauyin STEM, na wakilai, ko wanda ke ƙarƙashin bincike da yake tambaya daga ina ɗabi'ar samfurin ta fito: Nemotron ya cancanci bambancin.

10Tambayoyin da ake yawan yi

---

Majiyoyi: Rahoton fasaha na NVIDIA Nemotron 3 Ultra · Shafin masu haɓakawa na NVIDIA (Yuni 2026) · Bayanan gine-ginen Nemotron 3 Super · Llama-Nemotron: Efficient Reasoning Models · Cerebras Research, REAP. Ma'aunan tushen NVIDIA ne ya bayar da rahotonsu a kan samfurin da ba a matsa ba sai in an danganta su ga Artificial Analysis; ginin da aka matsa na iya bambanta. Farashi da iyawa na yanzu ne tun 5 ga Satumba, 2026.