I-API ye-DeepSeek Engavinjelwe

Abantu abafika emshweni othi "i-API ye-DeepSeek engavinjelwe" bafuna into eyodwa ethize: imodeli enekhono lokucabanga nokubhala ikhodi le-DeepSeek ephendula umbuzo abawubuze ngempela. Ingxenye yekhono ibhalwe kahle — i-DeepSeek-AI ishicilele ukwakhiwa namanani, futhi izisindo ziku-Hugging Face ngaphansi kwe-MIT. Ingxenye yesibili yisinqumo sokusebenza, futhi ngokuvamile senzelwa wena. Leli khasi lifaka kokubili: ukuthi siyini isisekelo se-V4-Pro, ukuthi ukwakhiwa kwethu kwe-3BIT-REAP kwenzani kuso, nokuthi u-"ongavinjelwe" usho ukuthini ncamashi lapho sikusho.

01Kusho ukuthini ngempela u-"i-API ye-DeepSeek engavinjelwe"

Izindlela ezimbili ezahlukene zifakwa ngaphansi kwegama elilodwa. Eyokuqala ukuqondaniswa okubhakwe ezisindweni — ukuziphatha kokwenqaba okufundwe ngesikhathi sokuqeqeshwa kwangemuva. Eyesibili ungqimba lokulawula okuqukethwe ohlangothini lokukhonza: isihlukanisi esizimele esihlola umyalelo wakho ngaphambi kokuba imodeli isibone kanye nokukhishwayo ngaphambi kokuba ukubone, futhi esingayinquma impendulo phakathi nomusho kungakhathalekile ukuthi imodeli ibifuna ukuthini.

Kokubili kuyalinganiseka, futhi insizakalo yokuqala ye-DeepSeek isilinganisiwe. I-Promptfoo yahlanganisa iqoqo lemiyalelo engu-1,360 yemibuzo ethinta ezombusazwe yabika ukuthi i-R1 ebiyisingathwe yi-DeepSeek yenqaba noma yagwema engu-1,156 yayo — cishe u-85%. Umsebenzi wezemfundo kulo mndeni ofanayo wathola indlela yokwehluleka ehlukile ababhali abayibiza ngokuthi ukucindezelwa kwemicabango, lapho umkhondo wokucabanga wemodeli uphela cishe uma nje uqala esihlokweni esibucayi. Ukuhlolwa okuhlukile okushicilelwe ku-Information Sciences kwabhala phansi ukucindezelwa ezingeni lencazelo: okuqukethwe okukhona ekucabangeni kwangaphakathi kwemodeli okususwa noma okubhalwa kabusha ngokuthula ngaphambi kokuba kufike kumsebenzisi. Leyo yinkinga enzima kunokwenqaba okusobala, ngoba impendulo enqunyiwe ibukeka njengempendulo.

Esikuqhubayo thina ukukhishwa kwezisindo ezivulekile, kuma-hardware esiwaqhubayo, e-United States, ngaphansi kwe-Shannon Lab LLC. Alukho ungqimba lokwenqaba phambi kwemodeli futhi asikho isihlungi esisetshenziswa kulokho okubuyayo: umyalelo wakho awuhlukaniselwa inqubomgomo ngaphambi kokuba ukhonzwe, futhi umsinga wokukhishwayo awuhlolwa endleleni ephuma ngayo. Ngenxa yokuthi isisekelo sinelayisensi ye-MIT futhi siyavela, ungaqhathanisa okukhishwayo kwethu nezisindo zereferensi ngokwakho kunokusithemba nje.

02Isisekelo: i-DeepSeek-V4-Pro, i-MoE engu-1.6T

I-DeepSeek-V4 wumndeni wamanje wezisindo ezivulekile we-DeepSeek-AI, owabonwa kuqala ngo-Ephreli 2026 wathunyelwa ngaphansi kwe-MIT ngosayizi ababili: V4-Flash (engu-284B isewonke, engu-13B esetshenziswayo) kanye ne-V4-Pro (engu-1.6T isewonke, engu-49B esetshenziswayo). Lesi sihloko simayelana nezinga le-Pro. Zombili zaqeqeshwa ngaphambili cishe ngamathokheni angu-33 trillion futhi zombili zisekela ngokwemvelo umongo wamathokheni ayisigidi.

Isilinganiso sokusabalala yikho okufanele ukuqaphele: cishe u-3% kuphela wamapharamitha osebenza kunoma yiliphi ithokheni, okuyikho okwenza imodeli enkulu kangaka ikwazi ukukhonzwa nhlobo. Ibhlogo ye-MoE uqobo isengu-DeepSeekMoE — ochwepheshe abaqondiswayo abancane kanye nochwepheshe abasetshenziswa njalo abethulwa ku-V2 — kunezinguquko ezincane kuphela ku-V4, okusobala kakhulu ukuqondiswa kwe-hash ezingqimbeni zokuqala. I-Multi-Token Prediction, inhloso esizayo ye-V3 evumela nokuqagela kokukhipha ikhodi, igcinwe ingashintshile.

Izingxenye ezintsha ku-V4 zingezokwakhiwa. I-Manifold-Constrained Hyper-Connections (mHC) ithatha indawo yendlela ejwayelekile yensalela ngemethriksi exubayo efundiwe evinjelwe ku-manifold wamamethriksi angokwenzeka kabili nge-Sinkhorn–Knopp projection — umsinga wensalela olinganisa kabusha ngokokufakwayo ngaphandle kokushelela kwezinombolo okubulala izitaki ezijule kakhulu. Ukuqeqeshwa kwasuka ku-AdamW kwaya ku-Muon ezingxenyeni eziningi, kusetshenziswa i-Newton–Schulz orthogonalization, kanti i-AdamW yagcinwa kuphela emabhalweni afakiwe, ezinhlokweni zokubikezela, ekulinganisweni nakumapharamitha ami e-mHC. Futhi izisindo zochwepheshe ziqeqeshwa ku-FP4 kanti amanye amapharamitha amaningi ase-FP8 — kucatshangelwa ukukalwa kwezinga kusukela ekuqaleni, okubalulekile ekutheni imodeli ikwamukela kanjani ukucindezelwa okwengeziwe.

03Kusuka ku-MLA kuya ku-CSA + HCA: indlela i-DeepSeek egcina ngayo umongo omude ushibhile

Umugqa oqhubekayo we-DeepSeek ezizukulwaneni ezine akusibo ubukhulu, kungumnotho wokunaka. Inguqulo ngayinye ihlasele i-KV cache, ekhula ngokulingana nomongo futhi ebusa inkumbulo ebudeni obude.

• I-V2 yethula i-Multi-head Latent Attention (MLA) — ukucindezelwa okuhlangene okusezingeni eliphansi kwezikhiye namanani kube yi-latent vector, ngakho ukukhiqiza kugcina leyo vekhtha kanye nokhiye we-RoPE ohlukanisiwe esikhundleni sokugcina u-K no-V bagcwele ngenhloko ngayinye. I-DeepSeek yabika i-KV cache yehla yaba cishe ngu-4–14% wokunaka okujwayelekile okunezinhloko eziningi, ikhwalithi igcinwa noma ithuthukiswa.
• I-V3 yakhulisa lokho yaya ku-671B, yenza i-MLA ibe ngumcabango ozenzakalelayo womongo omude oshibhile emamodelini avulekile.
• I-V3.2 yengeza i-DeepSeek Sparse Attention (DSA), ikhetha ingxenye ephezulu ye-top-k yomlando ngombuzo ngamunye ngenkomba elula kunokunaka konke.
• I-V4 iyaxuba. I-Compressed Sparse Attention (CSA) iqala icindezele okungenayo kwe-KV kwamathokheni ngalinye elingu-m kube yinto eyodwa ngezisindo ezifundiwe, bese iqhuba ukukhetha okusabalele kwe-top-k phezu kwalokho okucindeziwe. I-Heavily Compressed Attention (HCA) icindezela ngamandla kakhulu kodwa ihlale iminyene phezu komphumela. Lezi zinhlobo ezimbili zezingqimba ziyaxutshwa kuwo wonke umbhoshongo, ngakho isifinyezo esishibhile somhlaba wonke nokutholwa okunembayo kwendawo kuyakwazi ukuhlala ndawonye.

Inzuzo ebikiwe emongweni wamathokheni ayisigidi iyamangaza: i-V4-Pro idinga cishe u-27% wama-FLOP okuqagela ngethokheni no-10% we-KV cache ye-V3.2. Yingakho isicelo se-262K siziphatha njengesicelo esijwayelekile kunesigameko senkumbulo, futhi yingakho amanani okutholwa komongo omude ehlala emile kunokuwa ngemuva kwamashumi ambalwa ezinkulungwane zamathokheni.

04Kusho ukuthini isijobelelo esithi 3BIT-REAP

I-id yekhathalogu yethu iphetha ngo-3BIT-REAP, futhi lokho kuchaza izinyathelo ezimbili ezizimele zokucindezela ezisetshenziswe ngokulandelana.

I-REAP: ukuthena ochwepheshe, hhayi ukubahlanganisa

I-REAP yi-Router-weighted Expert Activation Pruning, evela ku-Cerebras Research — "REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression", eyamukelwa ku-ICLR 2026. Ku-MoE esabalele i-router ayibasebenzisi ngokulinganayo ochwepheshe bayo: umsila omude ukhethwa kancane, ngamandla amancane, noma kokubili. I-REAP ilinganisa uchwepheshe ngamunye ngokuhlanganisa amanani egeti ye-router (kangaki futhi ngamandla amangakanani ekhethwa) namazinga okusebenza kochwepheshe (kufaka kangakanani okukhishwayo kwakhe), bese ikhipha abehlulekile ngokuphelele.

Impikiswano yephepha ukuthi ukuthena kudlula ukuhlanganisa: ukuhlanganisa kususa ngesilinganiso ukukhethekile komsebenzi okwenza i-MoE ifanelwe ukwakhiwa, kanti ukukhipha ochwepheshe i-router engabafinyeleli kahle kuyakugcina. Emamodelini asuka ku-20B kuya ku-1T, i-REAP yadlula ukuhlanganisa nezinye izindlela zokuthena ekuhlolweni kokukhiqiza, futhi ababhali babika imiphumela ecishe ingalahli lutho ekukhiqizeni ikhodi nasekubizeni amathuluzi ekuthenweni kwabo-50% bochwepheshe ku-Qwen3-Coder-480B naku-Kimi-K2 — okuyiyona kanye imodi yomsebenzi le modeli ekhona ngenxa yayo.

Izisindo zamabhithi amathathu: izinga lokucindezela lezisekelo ezinkulu kunazo zonke

Lokho okusalayo bese kukalwa kube yizisindo zamabhithi amathathu. Amabhithi amathathu ayaqinisa, futhi asenzi sengathi akunjalo: ohlwini lonke sisebenzisa u-W8A16 lapho isisekelo sisincane ngokwanele ukuze singadingi ukucindezelwa, u-W4A16-AUTOROUND-REAP maphakathi, kanti u-3BIT-REAP sisebenzisa kuphela ezisekelweni ezinkulu kunazo zonke, lapho kungumehluko phakathi kokukhonza imodeli nokungayikhonzi.

Ukushintshana kufanele kushiwo ngokusobala: i-MoE engu-1.6T yamabhithi amathathu ayifani ngebhithi ngalinye nokukhishwa kwereferensi, futhi izindleko ezincane zekhwalithi zingezangempela. Okutholayo ngenxa yalokho yisisekelo esingacindezelwe esinophahla oluphezulu kakhulu kunanoma yini elingana kalula ngokunemba okuphezulu — futhi ngenxa yokuthi ochwepheshe be-V4 baqeqeshwa ku-FP4 kwasekuqaleni, imodeli iziphatha kahle ngokungajwayelekile ngaphansi kokucindezelwa okwengeziwe kwezisindo. Uma ufuna ukwakhiwa kwe-DeepSeek okunesandla esilula ekunembeni, u-DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP usebenzisa izisindo zamabhithi amane nokusebenza kwamabhithi ayi-16 kanye ne-AutoRound's learned rounding, kuwo umongo ofanayo we-262K ku-$0.50 ongenayo / $2.00 ophumayo.

05Ukucabanga, kanye nesishibilikisi se-reasoning_effort

Ukuqeqeshwa kwangemuva kwe-V4 kuqeqesha ochwepheshe bezindawo ngokwehlukene — izibalo, ukubhala ikhodi, ama-ejenti, ukulandela imiyalelo — nge-SFT elandelwa yi-GRPO, indlela ye-Group Relative Policy Optimization i-DeepSeek eyethula ngayo i-R1. Laba ochwepheshe bese behlanganiswa babe yimodeli eyodwa ngokuhlukaniswa okusenqubweni ngaphansi kwenhloso ye-reverse-KL, ngakho ikheshiphoyinti elilodwa liphethe ukuziphatha okusezingeni lochwepheshe ezindaweni eziningana ngesikhathi esisodwa kunokuba yimodeli evamile engabalulekile kuzo zonke.

Imodeli ekhishiwe iveza izimodi zokucabanga ezinamazinga kunenkinobho eyodwa yokuvala nokuvula. Kusango lethu leyo yipharamitha ejwayelekile ye-reasoning_effort, isishibilikisi esiwusizo kunazo zonke kule modeli:

• Phansi — ukuphendula okucishe kube okuqondile: ukuhlukanisa ngezigaba, ukukhipha ulwazi, ukubhala kabusha, noma yini lapho ukhokhela ulwazi lwemodeli kunokucabanga kwayo.
• Maphakathi — imodi yokusebenza ezenzakalelayo yokubuyekeza ikhodi, ukususa amaphutha nokuhlaziya okunezinyathelo eziningi.
• Phezulu — ukucabanga isikhathi eside ngaphambi kokuphendula. Lapho izinkinga ezinzima zixazululwa khona, nalapho ibhili yakho yamathokheni akhishwayo ihlala khona, ngoba amathokheni okucabanga akhokhiswa njengokukhishwayo.

Ngenxa yokuthi alukho ungqimba lokwenqaba, umkhondo wokucabanga awunqunywa ezihlokweni ezibucayi — iphethini yokucindezelwa kwemicabango ebhalwe phansi kunsizakalo yokuqala ayinayo indlela yokwenzeka lapha.

06Ikhodi nomsebenzi we-ejenti: okulinganiswe yisisekelo

Lawa ngamanani ashicilelwe yi-DeepSeek-AI uqobo okukhishwa kwe-V4-Pro — akusiwo awethu, futhi awalinganiselwanga ekwakhiweni kwethu okucindeziwe. Siwakhombisa ngoba ayisizathu esiqotho sokukhetha lesi sisekelo, futhi siwaphawula njengawenkampani yokuqala ukuze uwakale ngokufanele.

Amabili afanele ukuphawulwa. I-SWE-bench Verified ku-80.6% ingukuxazulula izinkinga zendawo yokugcina ikhodi yangempela — ukufunda ikhodi ongayijwayele, ukuthola iphutha, ukukhiqiza isichibiyelo esidlula ukuhlolwa kwephrojekthi ngokwayo. I-MRCR emathokhenini ayisigidi (83.5) ingukutholwa kwe-coreference kwemizuliswano eminingi ebudeni obugcwele, inombolo ekutshela ukuthi isimangalo somongo omude asisiso isihlobiso. I-DeepSeek ibika futhi isilinganiso se-Codeforces esingu-3206 kule modeli, kanye no-62.0 ku-CorpusQA emongweni wayi-1M.

Amanani angenhla abikwe yi-DeepSeek-AI okukhishwa kwereferensi kwe-DeepSeek-V4-Pro. Ukwakhiwa kwethu kwe-3BIT-REAP kuyinguqulo ecindeziwe yalezo zisindo futhi akukahlolwa ngokuzimele kule misebenzi; phatha amaphuzu ereferensi njengophahla lwesisekelo, hhayi njengokulinganiswa kwalesi siphetho.

07Umongo we-262K: yini engena ngempela

Imodeli yereferensi iqeqeshelwe amathokheni ayisigidi. Sikhonza angu-262,144 — iwindi yonke imodeli esangweni elabelana ngalo, ngakho ukushintsha amamodeli kungushintsho lomugqa owodwa futhi akulokothi kube ukwehla komongo. Empeleni u-262K udlulele ephuzwini lapho umongo ukhawulela khona umsebenzi omningi wangempela:

Umsebenzi	Cishe amathokheni	Kuyangena ku-262K?
Indawo yokugcina ikhodi ephakathi (~50k imigqa)	~600K	Cha — khetha izingxenye
Ingxenye eyodwa yohlelo nokuhlolwa kwayo	~60–120K	Yebo, kalula
Iqoqo eliphelele lama-RFC eprothokholi	~150K	Yebo
Incwadi yobuchwepheshe yamakhasi angu-700	~250K	Yebo, seduze nomkhawulo
Ukusebenza kwe-ejenti okude, imizuliswano yamathuluzi engu-40+	~80–200K	Yebo

Umkhawulo okufanele uwuhlelele ngokukhishwayo: amathokheni angu-16,384 aphezulu empendulweni ngayinye. Lokho kuyidokhumenti ende noma isichibiyelo esikhulu esinamafayela amaningi, kodwa akusiyo indawo yokugcina ikhodi. Imisebenzi emikhulu yokukhiqiza yenziwa ngokuxhuma izicelo ngokubuyisela okukhishwayo kwangaphambili emongweni — okwenziwa kushibhe yiwindi le-262K — kunokucela ukugcwaliswa okukodwa okukhulu.

08Intengo, nokuthi i-V4-Pro ihlalaphi ohlwini

Ku-$1.95 engenayo / $3.90 ephumayo ngethokheni eyisigidi, i-V4-Pro yizinga eliphezulu lokucabanga — imodeli oyiyisayo lapho inkinga inzima. Lapho inkulu nje, amamodeli ashibhile ewindini elifanayo le-262K azoyenza ngengxenye eyodwa kwezine yenani.

Imodeli engavinjelwe	Engenayo / 1M	Ephumayo / 1M	Umbono	Ilungele
DeepSeek-V4-Pro · 3BIT-REAP	$1.95	$3.90	Cha	Ukucabanga okunzima, ikhodi, ama-ejenti
DeepSeek-V4-Flash · W4A16	$0.50	$2.00	Cha	Uhlobo olufanayo, ivolumu ephezulu
GLM-5.2 · 3BIT-REAP	$0.73	$2.34	Cha	Inkabi yemisebenzi evamile
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	Yebo	Ama-ejenti aphezulu anezindlela eziningi
MiniMax-M3 · 3BIT-REAP	$0.50	$2.00	Yebo	Umbono oshibhe kunazo zonke

Sikhuluma iqiniso ngokuqhathaniswa abantu abazokwenza ngempela: i-API ye-DeepSeek uqobo ishibhile kunalena — i-V4-Pro yenkampani yokuqala ibhalwe cishe ku-$0.435 engenayo / $0.87 ephumayo, futhi intengo yabo yokufakwayo okutholakala kukheshi isashibhile nakakhulu. Uma intengo ngethokheni iyiyona kuphela into oyibhekayo futhi ukuziphatha kokulawula okuqukethwe kungawuthinti umsebenzi wakho, sebenzisa i-API yenkampani yokuqala; singathanda ukukutshela lokho kunokuba uzitholele ngokwakho kamuva. Inani elengeziwe lithenga iqoqo elehlukile lezimfanelo: alukho ungqimba lokwenqaba futhi asikho isihlungi sokukhishwayo, umqhubi wase-US ngaphansi kwe-Shannon Lab LLC, ikheshiphoyinti elibophezelwe elinosuku (0813) ukuze ukwenyuselwa okuthule kwangaphezulu kungalokothi kushintshe okukhishwayo kwakho phakathi kwephrojekthi, izilimi ezintathu ze-API kwimodeli eyodwa, kanye newindi elifanayo le-262K nendawo yamapharamitha efanayo kuwo wonke amamodeli ayishumi nambili ohlwini.

09Ukuyibiza: izilimi ezintathu, imodeli eyodwa

Yonke imodeli esangweni iphendula kuzo zontathu izilimi, ngokusakaza kuzo zonke. Awamukeli i-SDK entsha — ushintsha i-URL eyisisekelo kanye ne-id yemodeli.

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 — ulimi lwe-OpenAI; lusebenza nama-SDK e-OpenAI nanoma yini eyakhelwe kuwo.
• /v1/messages — ulimi lwe-Anthropic, lwezindawo zokugcina ikhodi esezibhalelwe leso simo.
• /v1/responses — isixhumanisi esisesimweni se-Responses.

Okusekelwayo kule modeli: ukusakaza, ukubiza amathuluzi kwemvelo, okukhishwayo okuhlelekile okuqinile kwe-json_schema, response_format, kanye no-reasoning_effort. Ireferensi ephelele yamapharamitha nendawo yokudlala ephilayo ziku-mibhalo ye-API.

10Imikhawulo eqotho

Izinto ezine okufanele uzazi ngaphambi kokwakha kuyo:

• Awukho umbono. Umbhalo kuphela; ngeke yamukele okufakwayo okuyisithombe kunoma yisiphi isiphetho. Uma udinga imodeli engavinjelwe ebonayo, sebenzisa i-Kimi-K3, i-Kimi-K2.6, i-MiniMax-M3, i-MiMo-V2.5 noma i-inkling.
• Uphahla lwamathokheni angu-16,384 okukhishwayo. Amathokheni okucabanga abalwa kulo emzamweni ophezulu, ngakho ukucabanga isikhathi eside kushiya indawo encane yempendulo. Xhuma izicelo emisebenzini emikhulu yokukhiqiza.
• Amabhithi amathathu ayikucindezelwa kwangempela. Amaphuzu ereferensi achaza isisekelo, hhayi lokhu kwakhiwa. Ngezinga eliphezulu kunazo zonke lokunemba, amamodeli e-W8A16 (MiMo-V2.5-Pro, MiMo-V2.5, Hy3) anokukalwa okuncane kunazo zonke.
• Okungavinjelwe akusho ukungabi nemiphumela. Ukungabikho kongqimba lokwenqaba kusho ukuthi isibopho ngalokho okubuzayo, nangalokho okwenzayo ngempendulo, sihlala kuwe — ngaphansi kwemigomo yethu yesevisi nomthetho wendawo osebenza kuyo.

11Imibuzo ebuzwa njalo

---

Imithombo: DeepSeek-AI — ikhadi lemodeli le-DeepSeek-V4-Pro · Umbiko wobuchwepheshe we-DeepSeek-V4 (CSA/HCA, mHC, Muon) · DeepSeek-V2 (MLA, DeepSeekMoE) · Umbiko wobuchwepheshe we-DeepSeek-V3 · REAP the Experts (Cerebras Research) · Ukuhlolwa kokuvinjelwa kwe-Promptfoo · Discovering Forbidden Topics in Language Models. Amanani okuhlolwa abikwe yi-DeepSeek-AI okukhishwa kwereferensi; ukuqinisekiswa okuzimele kwenkampani yesithathu kokwakhiwa okucindeziwe kusalindile. Amanani ayintengo eshicilelwe ye-Shannon AI kusukela ngoMandulo 5, 2026.