API na DeepSeek Marar Tantancewa

Mutanen da suke isowa ta kalmar "API na DeepSeek marar tantancewa" suna son abu guda takamaimai: samfuri mai iyawar tunani da rubuta lamba irin ta DeepSeek wanda yake amsa tambayar da suka yi da gaske. An rubuta rabin iyawar sosai — DeepSeek-AI ya wallafa gine-ginen da lambobin, kuma nauyin yana kan Hugging Face ƙarƙashin MIT. Rabi na biyu shawarar gudanarwa ce, kuma yawanci ana yanke ta a madadinka. Wannan shafin ya ƙunshi duka biyun: menene tushen V4-Pro, me ginin mu na 3BIT-REAP yake yi da shi, da kuma ainihin abin da "marar tantancewa" yake nufi sa'ilin da muka faɗe shi.

01Menene "API na DeepSeek marar tantancewa" da gaske

Hanyoyi biyu daban-daban ana sanya su ƙarƙashin kalma ɗaya. Ta farko ita ce daidaitawar da aka gasa cikin nauyin — ɗabi'ar ƙi da aka koya lokacin horo na baya. Ta biyu ita ce shimfidar tantancewa a gefen hidima: wani mai rarrabewa daban wanda yake duba umarninka kafin samfurin ya gan shi, yana kuma duba fitarwa kafin ka gan ta, kuma yana iya yanke amsa a tsakiyar jimla ba tare da la'akari da abin da samfurin ke son faɗa ba.

Ana iya auna duka biyun, kuma an auna hidimar DeepSeek ta asali. Promptfoo ya tara tambayoyi 1,360 masu taɓa siyasa sannan ya bayar da rahoton cewa R1 da DeepSeek ke ɗaukar nauyinsa ya ƙi ko ya karkatar da 1,156 daga cikinsu — kusan kashi 85%. Aikin ilimi a kan dangi ɗaya ya gano wata gazawa dabam da marubutan suka kira danne tunani, inda alamar tunanin samfurin take mutuwa kusan da zarar ta fara a kan batu mai laushi. Wani binciken daban da aka wallafa a Information Sciences ya rubuta danne matakin ma'ana: abun ciki da yake nan cikin tunanin samfurin na cikin gida amma a shiru ake barin sa ko a sake rubuta shi kafin ya isa ga mai amfani. Wannan ya fi ƙi bayyananne wuya, saboda amsa da aka yanke tana kama da amsa.

Abin da muke gudanarwa shi ne sakin buɗaɗɗen nauyi, a kan na'urorin da muke aiki da su, a cikin Amurka, ƙarƙashin Shannon Lab LLC. Babu shimfidar ƙi a gaban samfurin kuma babu matata a kan abin da ke dawowa: ba a rarraba umarninka bisa manufa kafin a yi masa hidima ba, kuma ba a duba yawan fitarwa a hanyarsa ta fita ba. Saboda tushen yana da lasisin MIT kuma na jama'a ne, kana iya kwatanta fitarwarmu da nauyin tunani da kanka maimakon ka amince da maganarmu kawai.

02Tushen: DeepSeek-V4-Pro, MoE mai 1.6T

DeepSeek-V4 shi ne dangin buɗaɗɗen nauyi na yanzu na DeepSeek-AI, an nuna shi a watan Afrilu 2026 kuma an fitar da shi ƙarƙashin MIT a girma biyu: V4-Flash (jimla 284B, 13B mai kunnuwa) da V4-Pro (jimla 1.6T, 49B mai kunnuwa). Wannan labarin game da matakin Pro ne. An riga an horar da duka biyun a kan kimanin tokens tiriliyan 33 kuma duka biyun suna goyon bayan mahalli na tokens miliyan tun asali.

Abin da ya kamata a lura da shi shi ne rabon tsattsauran aiki: kusan kashi 3% na sigogin ne kawai suke kunnuwa a kowane token, wanda shi ne ya sa samfuri mai wannan girman ya yiwu a yi masa hidima kwata-kwata. Toshen MoE da kansa har yanzu DeepSeekMoE ne — masanan da aka tura wa aiki a mataki mai kyau tare da masanan da ake raba wa aiki koyaushe, waɗanda aka gabatar a V2 — da ƙananan canje-canje kawai a V4, wanda ya fi bayyana shi ne turawa ta hash a matakan farko. An riƙe Multi-Token Prediction, manufar taimako ta V3 wadda take kuma ba da damar speculative decoding, ba tare da canji ba.

Sabbin sassan V4 na gine-gine ne. Manifold-Constrained Hyper-Connections (mHC) sun maye gurbin hanyar sauran mai sauƙi da matrix ɗin haɗawa da aka koya wanda aka takura shi a kan manifold na matrices masu tsattsauran biyu ta hanyar tsinkayar Sinkhorn–Knopp — rafin sauran wanda yake sake auna nauyi bisa kowace shigarwa ba tare da karkacewar lissafi da take kashe manyan tarin da suka yi zurfi ba. Horo ya bar AdamW zuwa Muon ga yawancin sassan, ta amfani da daidaita kusurwoyi na Newton–Schulz, an bar AdamW kawai ga embeddings, kawunan hasashe, daidaitawa da sigogin mHC marasa motsi. Kuma ana horar da nauyin masana a FP4 yayin da yawancin sauran sigogin suke a FP8 — an tsara shi da matsawa a rai tun farko, wanda yake da muhimmanci ga yadda samfurin yake karɓar ƙarin matsawa cikin sauƙi.

03Daga MLA zuwa CSA + HCA: yadda DeepSeek yake sa dogon mahalli ya zama mai rahusa

Zaren da ya ratsa tsararraki huɗu na DeepSeek ba girma ba ne, tattalin arzikin kulawa ne. Kowace sigar ta far wa ma'ajiyar KV, wadda take girma kai tsaye da mahalli kuma take mamaye ƙwaƙwalwa a dogayen tsawo.

• V2 ya gabatar da Multi-head Latent Attention (MLA) — matsi haɗe mai ƙaramin matsayi na maɓallai da ƙimomi zuwa vector ɗin ɓoye, don haka samarwa tana ajiye wannan vector ɗin tare da maɓallin RoPE da aka raba maimakon cikakken K da V na kowane kai. DeepSeek ya bayar da rahoton cewa ma'ajiyar KV ta sauka zuwa kusan kashi 4–14% na kulawa ta gargajiya mai kawuna da yawa, tare da riƙe inganci ko ma ingantawa.
• V3 ya faɗaɗa hakan zuwa 671B, ya sa MLA ta zama tsammani na asali don dogon mahalli mai rahusa a buɗaɗɗun samfura.
• V3.2 ya ƙara DeepSeek Sparse Attention (DSA), yana zaɓar rukunin top-k na tarihi ga kowace tambaya ta hanyar mai fihirisa mai sauƙi maimakon ya kula da komai.
• V4 ya haɗa su. Compressed Sparse Attention (CSA) tana fara matsa shigarwar KV na kowane tokens m zuwa shigarwa guda da nauyin da aka koya, sannan ta gudanar da zaɓen top-k mai tsattsauri a kan waɗannan shigarwar da aka matsa. Heavily Compressed Attention (HCA) tana matsawa da ƙarfi sosai amma tana zama mai kauri a kan sakamakon. Ana haɗa nau'ikan shimfidu biyun a cikin tarin, don haka taƙaitawa ta duniya mai rahusa da dawo da bayani na kusa mai daidaito suna zama tare.

Ribar da aka bayar da rahoto a mahalli na tokens miliyan tana da ban mamaki: V4-Pro yana buƙatar kusan kashi 27% na FLOPs na tunani a kowane token da kashi 10% na ma'ajiyar KV na V3.2. Shi ya sa buƙata ta 262K take nuna hali kamar buƙata ta yau da kullum maimakon wani lamari na ƙwaƙwalwa, kuma shi ya sa lambobin dawo da bayani a dogon mahalli suke ci gaba da tsayawa maimakon su durƙushe bayan dubbai kaɗan na tokens.

04Menene ma'anar ƙarin 3BIT-REAP

Alamar katalogin mu tana ƙarewa da 3BIT-REAP, kuma wannan yana kwatanta matakan matsawa masu zaman kansu guda biyu da aka yi a jere.

REAP: datse masana, ba haɗa su ba

REAP shi ne Router-weighted Expert Activation Pruning, daga Cerebras Research — "REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression", an karɓe shi zuwa ICLR 2026. A cikin MoE mai tsattsauri router ɗin ba ya amfani da masananshi daidai wa daida: ana zaɓar dogon jela da wuya, a rarrauna, ko duka biyun. REAP yana ba kowane masani maki ta haɗa ƙimomin ƙofar router (sau nawa kuma da wane ƙarfi ake zaɓensa) da ma'aunin kunnawar masani (nawa fitarwarsa take bayarwa), sannan ya goge waɗanda suka faɗi kai tsaye.

Hujjar takardar ita ce datsawa ta fi haɗawa: haɗawa tana share ƙwarewar aiki ta musamman da ta sa MoE ta cancanci ginawa, alhali goge masanan da router bai kai gare su sosai ba yana kiyaye ta. A kan samfura daga 20B zuwa 1T, REAP ya zarce haɗawa da sauran hanyoyin datsawa a kan ma'aunan samarwa, kuma marubutan sun bayar da rahoton sakamako kusan mara asara a kan samar da lamba da kiran kayan aiki a datsa masana kashi 50% a kan Qwen3-Coder-480B da Kimi-K2 — wanda daidai shi ne bayanin nau'in aikin da wannan samfurin yake nan dominsa.

Nauyi na 3-bit: matakin matsawa ga manyan tushe

Abin da ya tsira sai a matsa shi zuwa nauyi na 3-bit. Bit uku yana da tsauri, kuma ba ma ɓoye hakan: a cikin jerin muna amfani da W8A16 inda tushen ya yi ƙanƙanta har ba ya buƙatar matsi, W4A16-AUTOROUND-REAP a tsakiya, da 3BIT-REAP a kan manyan tushe kawai, inda shi ne bambanci tsakanin yin hidimar samfurin da rashin yin hidimarsa.

Ya kamata a faɗi cinikin a fili: MoE na 1.6T a 3-bit ba iri ɗaya ba ne bit-da-bit da sakin tunani, kuma ƙaramar asarar inganci gaskiya ce. Abin da ka samu a maimakon shi ne tushe wanda rufinsa marar matsawa yake sama sosai da duk abin da zai shiga cikin sauƙi a daidaito mafi girma — kuma saboda an horar da masanan V4 a FP4 tun farko, samfurin yana nuna hali mai kyau na musamman a ƙarƙashin ƙarin matsa nauyi. Idan kana son ginin DeepSeek mai taɓawa mai laushi a kan daidaito, DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP yana amfani da nauyi na 4-bit da kunnuwa na 16-bit da zagayen da AutoRound ya koya, a kan mahalli 262K guda a $0.50 shiga / $2.00 fita.

05Tunani, da maɓallin reasoning_effort

Horon baya na V4 yana horar da ƙwararrun fannoni daban-daban — lissafi, rubuta lamba, wakilai, bin umarni — da SFT sannan GRPO, hanyar Group Relative Policy Optimization da DeepSeek ya gabatar da R1. Sannan ana haɗa waɗannan ƙwararrun zuwa samfuri guda ta hanyar tacewar ilimi bisa manufa ƙarƙashin manufar reverse-KL, don haka checkpoint guda yana ɗauke da ɗabi'a ta matakin ƙwararre a fannoni da dama a lokaci ɗaya maimakon ya zama mai kowa da kowa da yake matsakaici a kowane fanni.

Samfurin da aka fitar yana nuna hanyoyin tunani a matakai maimakon maɓalli na eh-ko-a'a. A ƙofar shigarmu wannan shi ne sigar reasoning_effort ta yau da kullum, maɓalli guda mafi amfani a kan wannan samfurin:

• Ƙasa — amsa kusan kai tsaye: rarrabewa, cirar bayani, sake rubutawa, duk inda kake biyan kuɗin ilimin samfurin maimakon nazarinsa.
• Tsakiya — yanayin aiki na asali don duba lamba, gyara kurakurai da nazari mai matakai da yawa.
• Sama — dogon nazari kafin amsawa. Inda ake warware matsaloli masu wuya, kuma inda kuɗin tokens ɗinka na fitarwa yake zama, tun da ana lissafin tokens na tunani a matsayin fitarwa.

Saboda babu shimfidar ƙi, ba a yanke alamar tunani a kan batutuwa masu laushi ba — tsarin danne tunani da aka rubuta a kan hidimar asali ba shi da wata hanyar da zai faru a nan.

06Lamba da aikin wakilai: makin da tushen ya samu

Waɗannan su ne lambobin da DeepSeek-AI da kansa ya wallafa don sakin V4-Pro — ba namu ba, kuma ba a auna su a kan ginin mu da aka matsa ba. Muna nuna su saboda su ne dalilin gaskiya na zaɓar wannan tushen, kuma muna alamta su a matsayin na asali domin ka auna su yadda ya dace.

Biyu sun cancanci sharhi. SWE-bench Verified a kashi 80.6% shi ne warware matsala a ainihin ma'ajiyar lamba — karanta tarin lamba da ba a saba da shi ba, gano aibin, samar da faci da zai ci jarabawar aikin da kansa. MRCR a tokens miliyan ɗaya (83.5) shi ne dawo da bayani na maimaita ambato a zagaye da yawa a cikakken tsawo, lambar da take gaya maka cewa iƙirarin dogon mahalli ba ado ba ne. DeepSeek ya kuma bayar da rahoton ƙimar Codeforces ta 3206 ga samfurin, da 62.0 a CorpusQA a mahalli na 1M.

DeepSeek-AI ne ya bayar da rahoton lambobin da ke sama ga sakin tunani na DeepSeek-V4-Pro. Ginin mu na 3BIT-REAP wani nau'i ne da aka matsa na wannan nauyin kuma ba a auna shi da kansa a waɗannan ayyukan ba; ka ɗauki makin tunani a matsayin rufin tushen, ba a matsayin auna wannan ƙarshen hanyar ba.

07Mahalli 262K: abin da yake shiga da gaske

An horar da samfurin tunani don tokens miliyan. Muna hidimar 262,144 — tagar da kowane samfuri a ƙofar shiga yake raba, don haka canja samfuri canjin layi guda ne kuma ba zai taɓa zama koma-baya na mahalli ba. A aikace 262K ya wuce inda mahalli yake ɗaure ga yawancin ainihin aiki:

Nau'in aiki	Kimanin tokens	Yana shiga cikin 262K?
Ma'ajiyar hidima matsakaiciya (~layuka 50k)	~600K	A'a — zaɓi sassa
Ƙaramin tsari guda da jarabawowinsa	~60–120K	Eh, cikin sauƙi
Cikakken jerin RFC na wata yarjejeniya	~150K	Eh
Littafin fasaha mai shafuka 700	~250K	Eh, kusa da iyaka
Dogon gudun wakili, zagayen kayan aiki 40+	~80–200K	Eh

Iyakar da za a yi shiri a kanta ita ce fitarwa: matsakaicin tokens 16,384 a kowace amsa. Wannan dogon takarda ne ko faci mai fayiloli da yawa mai muhimmanci, amma ba tarin lamba ba. Ana yin manyan ayyukan samarwa ta haɗa buƙatu a jere tare da mayar da fitarwar da ta gabata cikin mahalli — abin da tagar 262K take sa ya zama mai rahusa — maimakon ta neman kammalawa guda babba.

08Farashi, da inda V4-Pro yake a cikin jerin

A $1.95 shiga / $3.90 fita kowane tokens miliyan, V4-Pro shi ne babban matakin tunani — samfurin da za ka kai gare shi sa'ilin da matsalar take da wuya. Sa'ilin da take babba kawai, samfura masu rahusa a kan tagar 262K guda za su yi ta a kan kashi ɗaya bisa huɗu na farashin.

Samfuri marar tantancewa	Shiga / 1M	Fita / 1M	Gani	Ya fi dacewa da
DeepSeek-V4-Pro · 3BIT-REAP	$1.95	$3.90	A'a	Tunani mai wuya, lamba, wakilai
DeepSeek-V4-Flash · W4A16	$0.50	$2.00	A'a	Zuriya ɗaya, girma mai yawa
GLM-5.2 · 3BIT-REAP	$0.73	$2.34	A'a	Doki na aiki na kowane fanni
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	Eh	Manyan wakilai masu hanyoyi da yawa
MiniMax-M3 · 3BIT-REAP	$0.50	$2.00	Eh	Gani mafi rahusa

A yi gaskiya game da kwatancin da mutane za su yi da gaske: API na DeepSeek da kansa ya fi wannan rahusa — V4-Pro na asali yana jera kusan $0.435 shiga / $0.87 fita, kuma farashin shigarwarsu na cache-hit ma ya fi ƙasa. Idan farashi a kowane token shi ne ma'aunin ka kaɗai kuma ɗabi'ar tantancewa ba ta shafar aikinka ba, yi amfani da API na asali; mun fi so mu gaya maka hakan maimakon ka gano daga baya. Ƙarin kuɗin yana sayen wasu halaye daban: babu shimfidar ƙi kuma babu matatar fitarwa, mai gudanarwa a Amurka ƙarƙashin Shannon Lab LLC, checkpoint mai kwanan wata da aka ɗaure (0813) don haka haɓaka a shiru daga sama ba zai taɓa canja fitarwarka a tsakiyar aiki ba, yaruka uku na API a kan samfuri guda, da tagar 262K guda da sigogi iri ɗaya a duk samfura goma sha biyu na jerin.

09Kiransa: yaruka uku, samfuri guda

Kowane samfuri a ƙofar shiga yana amsawa a kan dukkan ƙarshen hanyoyi uku, da streaming a kowanne. Ba ka ɗauki sabon SDK ba — kana canja base URL da alamar samfuri kawai.

curl https://api.shannon-ai.com/v1/chat/completions \
  -H "Authorization: Bearer $SHANNON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "DeepSeek-V4-Pro-0813-3BIT-REAP",
    "reasoning_effort": "high",
    "stream": true,
    "messages": [
      {"role": "system", "content": "You are a systems programmer. Be direct."},
      {"role": "user", "content": "Audit this parser for memory-safety bugs."}
    ]
  }'

• /v1/chat/completions — yaren OpenAI; yana aiki da SDK na OpenAI da duk abin da aka gina a kansu.
• /v1/messages — yaren Anthropic, ga tarin lamba da aka riga aka rubuta bisa wannan siffar.
• /v1/responses — mahaɗar salon Responses.

Abin da ake goyon baya a kan wannan samfurin: streaming, kiran kayan aiki na asali, fitarwa mai tsari ta json_schema mai tsauri, response_format, da reasoning_effort. Cikakken bayanin sigogi da filin gwaji kai tsaye suna cikin takardun API.

10Iyakoki na gaskiya

Abubuwa huɗu da ya kamata ka sani kafin ka gina a kansa:

• Babu gani. Rubutu kaɗai; ba zai karɓi shigar hoto a kan kowane ƙarshen hanya ba. Idan kana buƙatar samfuri marar tantancewa da yake gani, yi amfani da Kimi-K3, Kimi-K2.6, MiniMax-M3, MiMo-V2.5 ko inkling.
• Rufin fitarwa na tokens 16,384. Tokens na tunani suna lissafuwa cikinsa a ƙoƙari mai girma, don haka dogon nazari yana barin ƙaramin wuri ga amsar. Ka haɗa buƙatu a jere don manyan ayyukan samarwa.
• 3-bit matsawa ce ta gaske. Makin tunani yana kwatanta tushen, ba wannan ginin ba. Ga matakin mafi aminci, samfuran W8A16 (MiMo-V2.5-Pro, MiMo-V2.5, Hy3) suna ɗauke da matsawa mafi ƙaranci.
• Rashin tantancewa ba ya nufin babu sakamako. Rashin shimfidar ƙi yana nufin alhakin abin da ka nema, da abin da ka yi da amsar, yana kanka — ƙarƙashin sharuɗɗan hidimarmu da dokar wurin da kake aiki.

11Tambayoyin da ake yawan yi

---

Majiyoyi: DeepSeek-AI — katin samfuri na DeepSeek-V4-Pro · Rahoton fasaha na DeepSeek-V4 (CSA/HCA, mHC, Muon) · DeepSeek-V2 (MLA, DeepSeekMoE) · Rahoton fasaha na DeepSeek-V3 · REAP the Experts (Cerebras Research) · Binciken tantancewa na Promptfoo · Discovering Forbidden Topics in Language Models. DeepSeek-AI ne ya bayar da rahoton lambobin ma'auni don sakin tunani; ana jiran tabbatarwa mai zaman kanta ta ɓangare na uku ga ginin da aka matsa. Farashin su ne farashin jerin na Shannon AI tun 5 ga Satumba, 2026.