API ya DeepSeek Yosasidwa

Anthu obwera pa mawu akuti "API ya DeepSeek yosasidwa" akufuna chinthu chimodzi enieni: mtundu wokhala ndi mphamvu ya DeepSeek yokulingalira ndi kulemba pulogalamu umene umayankha funso limene analifunsadi. Theka la mphamvuyo lafotokozedwa bwino — DeepSeek-AI idasindikiza kapangidwe ndi manambala, ndipo zolemera zili pa Hugging Face pansi pa MIT. Theka lachiwiri ndi chisankho cha kayendetsedwe, ndipo nthawi zambiri chimasankhidwa kwa inu. Tsamba ili likukhudza zonse ziwiri: chimene maziko a V4-Pro ali, chimene kapangidwe kathu ka 3BIT-REAP kamachita nawo, ndi chimene "chosasidwa" chikutanthauza kwenikweni pamene tikunena.

01Chimene "API ya DeepSeek yosasidwa" chikutanthauzadi

Njira ziwiri zosiyana zimaikidwa pansi pa liwu limodzi. Yoyamba ndi kugwirizanitsidwa kophikidwa mu zolemera zenizeni — khalidwe lokana lophunzitsidwa pa nthawi ya post-training. Yachiwiri ndi wosanjikiza wowunika wa kutumikira: chosankha chinthu chodziyimira pawokha chimene chimayang'ana prompt yanu mtunduwo usanaione ndi zotulutsa musanaziwone, ndipo chingadule yankho pakati pa chiganizo mosasamala kanthu chimene mtunduwo unkafuna kunena.

Zonse ziwiri zingayezedwe, ndipo ntchito ya DeepSeek yeniyeni yayezedwa. Promptfoo idasonkhanitsa mafunso 1,360 okhudza ndale zovuta ndipo idanena kuti R1 ya DeepSeek yotumizidwa idakana kapena kupatutsa 1,156 mwa mafunsowo — pafupifupi 85%. Ntchito ya akatswiri pa banja lomwelo idapeza njira yolakwika yosiyana imene olemba adaitcha thought suppression (kuponderezedwa kwa maganizo), pamene njira ya kulingalira ya mtunduwo imatha pafupifupi nthawi yomweyo pamene ikuyamba pa mutu wovuta. Kufufuza kwina kosindikizidwa mu Information Sciences kudalemba kuponderezedwa pa mulingo wa tanthauzo: zinthu zomwe zilipo mu kulingalira kwa mkati mwa mtunduwo zomwe zimasiyidwa mwakachetechete kapena kulembedwanso musanafike kwa wogwiritsa ntchito. Ndi vuto lovuta kuposa kukana konse, chifukwa yankho lodulidwa limawoneka ngati yankho.

Chimene timayendetsa ndi kutulutsidwa kwa zolemera zapoyera, pa zida zathuzathu, ku United States, pansi pa Shannon Lab LLC. Palibe wosanjikiza wa kukana patsogolo pa mtunduwo ndipo palibe chosefera chogwiritsidwa ntchito pa chimene chikubwerera: prompt yanu siikuikidwa m'gulu la ndondomeko musanaitumikire, ndipo zotulutsa sizikuyang'aniridwa pamene zikutuluka. Chifukwa maziko ake ali achilolezo cha MIT ndipo apoyera, mungathe kuyerekeza zotulutsa zathu ndi zolemera zoyambirira nokha m'malo motenga mawu athu okha.

02Maziko: DeepSeek-V4-Pro, MoE ya 1.6T

DeepSeek-V4 ndi banja lapano la zolemera zapoyera la DeepSeek-AI, lomwe linayamba kuonetsedwa mu Epulo 2026 ndipo linatulutsidwa pansi pa MIT mu makulidwe awiri: V4-Flash (284B onse, 13B ogwira ntchito) ndi V4-Pro (1.6T onse, 49B ogwira ntchito). Nkhaniyi ikukhudza mlingo wa Pro. Onse awiri anaphunzitsidwa poyamba pa pafupifupi ma token thililiyoni 33 ndipo onse awiri amathandizira mwachibadwa context ya ma token miliyoni imodzi.

Chiyerekezo cha kuchepa ndicho chinthu choyenera kudziwika: pafupifupi 3% chokha cha ma parameter amagwira ntchito pa token iliyonse, ndicho chimene chimapangitsa kuti mtundu waukulu ngati uwu uzitheke kutumikiridwa. Chigawo cha MoE chokha chikadali DeepSeekMoE — akatswiri oyendetsedwa amtundu wapamwamba kuphatikiza akatswiri ogawana amene amakhala pantchito nthawi zonse, oyambitsidwa mu V2 — ndi kusintha kwaking'ono kokha kwa V4, koonekera kwambiri ngati hash routing mu magawo oyambirira kwambiri. Multi-Token Prediction, cholinga chothandiza cha V3 chimene chimathandizanso speculative decoding, chasungidwa mosasintha.

Zigawo zatsopano mu V4 ndi zamapangidwe. Manifold-Constrained Hyper-Connections (mHC) zimasintha njira yosalala ya residual ndi matrix yophunzitsidwa yosakaniza yoletsedwa pa manifold ya matrix za doubly stochastic kudzera mu Sinkhorn–Knopp projection — mtsinje wa residual umene umasinthanso kulemera pa zolowetsa zilizonse popanda kupatuka kwa manambala kumene kumapha zomangamanga zakuya kwambiri. Kuphunzitsa kunachoka pa AdamW kupita ku Muon pa magawo ambiri, pogwiritsa ntchito Newton–Schulz orthogonalization, ndi AdamW yosungidwa pa embeddings, prediction heads, normalization ndi ma parameter okhazikika a mHC okha. Ndipo zolemera za akatswiri zimaphunzitsidwa mu FP4 pamene ma parameter ena ambiri ali mu FP8 — odziwa za quantization kuyambira pachiyambi, chomwe ndi chofunika pa mmene mtunduwo umalandirira kufinyidwa kowonjezera mosavuta.

03Kuchokera ku MLA kupita ku CSA + HCA: mmene DeepSeek imasungira context yayitali yotsika mtengo

Mzere waukulu wa DeepSeek pakati pa mibadwo inayi si kukula, ndi chuma cha attention. Mtundu uliwonse wathira nkhondo ku KV cache, imene imakula mwaulinganiro ndi context ndipo imalamulira kukumbukira pa kutalika kwakukulu.

• V2 idayambitsa Multi-head Latent Attention (MLA) — kufinya kwa keys ndi values kukhala vector imodzi ya latent yokhala ndi rank yotsika, choncho kupanga kumasunga vector imeneyo kuphatikiza decoupled RoPE key m'malo mwa K ndi V zathunthu pa head iliyonse. DeepSeek idanena kuti KV cache idatsika mpaka pafupifupi 4–14% ya multi-head attention wamba, ndi mtundu wosungidwa kapena wokwezedwa.
• V3 idakulitsa izi mpaka 671B, kupangitsa MLA kukhala chiyembekezo chachikale cha context yayitali yotsika mtengo mu mitundu yapoyera.
• V3.2 idawonjezera DeepSeek Sparse Attention (DSA), kusankha gawo la top-k la mbiri pa funso lililonse kudzera mu indexer yopepuka m'malo mongoyang'ana zonse.
• V4 imasakaniza njira ziwiri. Compressed Sparse Attention (CSA) poyamba imafinya zolowetsa za KV za token iliyonse ya m kukhala imodzi ndi zolemera zophunzitsidwa, kenako imayendetsa kusankha kwa top-k pa zolowetsa zofinyidwazo. Heavily Compressed Attention (HCA) imafinya mokulirapo koma imakhalabe dense pa chotulukacho. Mitundu iwiri ya magawo imasakanizidwa kudutsa zomangamanga zonse, choncho chidule chapadziko lonse chotsika mtengo ndi kutenga koyenera kwa malo kumakhala pamodzi.

Phindu lofotokozedwa pa context ya ma token miliyoni imodzi ndi lodziwikiratu: V4-Pro imafuna pafupifupi 27% ya FLOPs za inference pa token iliyonse ndi 10% ya KV cache ya V3.2. Ndicho chifukwa chake pempho la 262K limakhala ngati pempho wamba m'malo mochitika chokhudza kukumbukira, ndi chifukwa chake manambala a kutenga kwa context yayitali amakhalabe olimba m'malo mogwa pambuyo pa ma token masauzande angapo.

04Chimene chotsatira cha 3BIT-REAP chikutanthauza

Id yathu ya katalogi imatha ndi 3BIT-REAP, ndipo imeneyi imafotokoza masitepe awiri odziyimira pawokha a kufinya ogwiritsidwa ntchito motsatana.

REAP: kudula akatswiri, osati kuwaphatikiza

REAP ndi Router-weighted Expert Activation Pruning, kuchokera ku Cerebras Research — "REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression", yomwe idavomerezedwa ku ICLR 2026. Mu MoE ya sparse, router siigwiritsa ntchito akatswiri ake mofanana: mchira wautali umasankhidwa mosowa, mofooka, kapena zonse ziwiri. REAP imayeza katswiri aliyense mwa kuphatikiza router gate values (kuchuluka ndi mphamvu ya kusankhidwa kwake) ndi expert activation norms (kuchuluka kwa zotulutsa zake), kenako imachotsa opanda phindu kotheratu.

Mfundo ya pepalali ndikuti kudula kumaposa kuphatikiza: kuphatikiza kumachotsa mwa kuwerengera pakati kusankhika kwa ntchito kumene kunapangitsa MoE kukhala yoyenera kumangidwa, pamene kuchotsa akatswiri amene router siifika nawo pafupi kumasunga kusankhikako. Pakati pa mitundu kuchokera ku 20B mpaka 1T, REAP idaposa kuphatikiza ndi njira zina zodulira pa mayeso opanga, ndipo olembawo akunena zotsatira pafupifupi zopanda kutayika pa kupanga ma kodi ndi kuyitanitsa zida pa kudula kwa akatswiri kwa 50% pa Qwen3-Coder-480B ndi Kimi-K2 — chomwe ndicho chizindikiro cha ntchito imene mtunduwu ulipo kuchita.

Zolemera za 3-bit: mlingo wa kufinya wa maziko akulu kwambiri

Chimene chatsalira chimasandulidwa kukhala zolemera za 3-bit. Ma bit atatu ndi olimba, ndipo sitinamanyeko china: pa mndandanda wathu wonse timagwiritsa ntchito W8A16 kumene maziko ndi ang'ono mokwanira osafunikira kufinyidwa, W4A16-AUTOROUND-REAP pakati, ndi 3BIT-REAP pa maziko akulu kwambiri okha, kumene ndicho kusiyana pakati pa kutumikira mtunduwo ndi kusawutumikira.

Kusinthana kwake n'koyenera kunenedwa momveka: MoE ya 3-bit ya 1.6T siyofanana ndi bit ndi kutulutsidwa koyambirira, ndipo mtengo waung'ono wa mtundu ndi wenieni. Chimene mumapeza ndi maziko amene malire awo osafinyidwa ali kutali kwambiri ndi chilichonse chomwe chimakwana bwino pa kulondola kwapamwamba — ndipo chifukwa akatswiri a V4 anaphunzitsidwa mu FP4 kuyambira pachiyambi, mtunduwo umakhala wokhala bwino mosadziwikiratu pansi pa kufinyidwa kwina kwa zolemera. Ngati mukufuna kapangidwe ka DeepSeek kokhala ndi kukhudza kopepuka pa kulondola, DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP imagwiritsa ntchito zolemera za 4-bit ndi zogwira ntchito za 16-bit ndi kuzungulira kophunzitsidwa kwa AutoRound, pa context yomweyo ya 262K pa $0.50 kulowa / $2.00 kutuluka.

05Kulingalira, ndi chosinthira cha reasoning_effort

Post-training ya V4 imaphunzitsa akatswiri a magawo padera padera — masamu, kulemba pulogalamu, ma agent, kutsatira malangizo — ndi SFT yotsatiridwa ndi GRPO, njira ya Group Relative Policy Optimization imene DeepSeek idayambitsa ndi R1. Akatswiriwo amaphatikizidwa kukhala mtundu umodzi kudzera mu on-policy distillation pansi pa cholinga cha reverse-KL, choncho checkpoint imodzi imanyamula khalidwe la mlingo wa katswiri m'magawo angapo nthawi imodzi m'malo mokhala wodziwa zambiri koma wapakati m'zonse.

Mtundu wotulutsidwawo umasonyeza njira zolingalira zokwezekakweza m'malo mwa chosinthira cha ziwiri chokha. Pa gateway yathu ndicho parameter wamba wa reasoning_effort, chosinthira chothandiza kwambiri pa mtunduwu:

• Low — kuyankha pafupifupi mwachindunji: kuika m'magulu, kutulutsa, kulemba mwatsopano, chilichonse chimene mukulipirira chidziwitso cha mtunduwo m'malo mokulingalira kwake.
• Medium — njira yachikale yogwirira ntchito ya kuunika ma kodi, kupeza zolakwika ndi kusanthula kwa masitepe angapo.
• High — kulingalira kwautali musanayankhe. Pamene mavuto ovuta amathetsedwa, ndi pamene ndalama zanu za ma token otulutsa zimakhala, popeza ma token a kulingalira amalipiritsidwa ngati zotulutsa.

Chifukwa palibe wosanjikiza wa kukana, njira ya kulingalira siyidulidwa pa mitu yovuta — mtundu wa thought suppression wolembedwa pa ntchito yeniyeni ilibe njira yoti uchitikire pano.

06Ntchito za kodi ndi za agent: zimene maziko amapeza

Izi ndi manambala osindikizidwa a DeepSeek-AI okha a kutulutsidwa kwa V4-Pro — osati athu, ndipo sanayezedwe pa kapangidwe kathu ka quantized. Tikuwasonyeza chifukwa ndicho chifukwa choona chosankhira maziko amenewa, ndipo tikuwazindikiritsa ngati a kampani yoyambirira kuti muwayeze molingana.

Ziwiri ziyenera kunenedwapo. SWE-bench Verified pa 80.6% ndi kuthetsa mavuto a repository enieni — kuwerenga codebase yosadziwika, kupeza cholakwika, kupanga patch imene imadutsa mayeso a polojekitiyo okha. MRCR pa ma token miliyoni imodzi (83.5) ndi kutenga kwa coreference ya masitepe ambiri pa kutalika konse, nambala imene ikukuuzani kuti chonena cha context yayitali si chokongoletsera chabe. DeepSeek imanenanso Codeforces rating ya 3206 ya mtunduwo, ndi 62.0 pa CorpusQA pa context ya 1M.

Manambala pamwambapa ndi onenedwa ndi DeepSeek-AI a kutulutsidwa koyambirira kwa DeepSeek-V4-Pro. Kapangidwe kathu ka 3BIT-REAP ndi mtundu wofinyidwa wa zolemera zimenezo ndipo sikanayezedwe modziyimira pawokha pa ntchito izi; onani manambala oyambirirawo ngati malire apamwamba a maziko, osati muyeso wa endpoint iyi.

07Context ya 262K: chimene chimakwaniradi

Mtundu woyambirira waphunzitsidwa pa ma token miliyoni imodzi. Timatumikira 262,144 — zenera limene mitundu yonse pa gateway imagawana, choncho kusintha mtundu ndi kusintha kwa mzere umodzi ndipo sikubwerera m'mbuyo pa context. Pa ntchito zenizeni, 262K ndi kupitilira pamalire pamene context imalepheretsa ntchito zambiri zenizeni:

Ntchito	Ma token pafupifupi	Zikukwana mu 262K?
Repo ya pakati ya ntchito (~mizere 50k)	~600K	Ayi — sankhani magawo
Gawo limodzi la dongosolo + mayeso ake	~60–120K	Inde, mosavuta
Mndandanda wonse wa RFC wa protocol	~150K	Inde
Buku la ukadaulo la masamba 700	~250K	Inde, pafupi ndi malire
Kuyenda kwautali kwa agent, kuposa mabwerezedwe 40 a zida	~80–200K	Inde

Malire oyenera kukonzekera ndi zotulutsa: ma token 16,384 apamwamba pa yankho lililonse. Ndi chikalata chachitali kapena patch yaikulu ya mafayilo ambiri, koma osati codebase. Ntchito zazikulu za kupanga zimachitika mwa kulumikiza mapempho ndi zotulutsa zapitazo m'context — chimene zenera la 262K chimapangitsa kukhala chotsika mtengo — m'malo mopempha zotulutsa zazikulu imodzi.

08Mtengo, ndi pamene V4-Pro ili mu mndandanda

Pa $1.95 kulowa / $3.90 kutuluka pa ma token miliyoni, V4-Pro ndi mlingo wapamwamba kwambiri wa kulingalira — mtundu umene mumafunafuna pamene vutolo ndi lovuta. Pamene ndi lalikulu lokha, mitundu yotsika mtengo pa zenera lomwelo la 262K idzachita pa gawo limodzi mwa magawo anayi a mtengo.

Mtundu wosasidwa	Kulowa / 1M	Kutuluka / 1M	Masomphenya	Wabwino kwambiri pa
DeepSeek-V4-Pro · 3BIT-REAP	$1.95	$3.90	Ayi	Kulingalira kovuta, ma kodi, ma agent
DeepSeek-V4-Flash · W4A16	$0.50	$2.00	Ayi	Mzera womwewo, kuchuluka kwakukulu
GLM-5.2 · 3BIT-REAP	$0.73	$2.34	Ayi	Wogwira ntchito zamba zambiri
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	Inde	Ma agent apamwamba a multimodal
MiniMax-M3 · 3BIT-REAP	$0.50	$2.00	Inde	Masomphenya otsika mtengo kwambiri

Kunena zoona pa kuyerekeza kumene anthu adzachite kwenikweni: API ya DeepSeek yokha ndi yotsika mtengo kuposa iyi — V4-Pro yeniyeni imalembedwa pafupifupi $0.435 kulowa / $0.87 kutuluka, ndipo mtengo wawo wa cache-hit ndi wotsikirapo. Ngati mtengo pa token ndicho chokhacho chimene mukuganizira ndipo khalidwe la kuwunika silikhudza ntchito yanu, gwiritsani ntchito API yeniyeniyo; tikanakonda kukuuzani izi kuposa kuti mudzapeze pambuyo pake. Mtengo wapamwambawu umakupatsani zinthu zosiyana: palibe wosanjikiza wa kukana ndipo palibe chosefera cha zotulutsa, woyendetsa wa US pansi pa Shannon Lab LLC, checkpoint yokhazikika ya tsiku (0813) kuti kukweza kwakachetechete kwa upstream kusasinthe zotulutsa zanu pakati pa polojekiti, zilankhulo zitatu za API pa mtundu umodzi, ndi zenera lomwelo la 262K ndi parameter surface pa mitundu yonse khumi ndi iwiri mu mndandanda.

09Kuyitana: zilankhulo zitatu, mtundu umodzi

Mtundu uliwonse pa gateway umayankha pa ma endpoint onse atatu, ndi streaming pa iliyonse. Simufunika kutenga SDK yatsopano — mumangosintha base URL ndi id ya mtundu.

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 — chilankhulo cha OpenAI; chimagwira ntchito ndi ma SDK a OpenAI ndi chilichonse chomangidwa pa iwo.
• /v1/messages — chilankhulo cha Anthropic, kwa ma codebase omwe analembedwa kale motsatira mawonekedwe amenewo.
• /v1/responses — mawonekedwe a Responses-style.

Zothandizidwa pa mtunduwu: streaming, kuyitanitsa zida mwachibadwa, zotulutsa zolimba za json_schema, response_format, ndi reasoning_effort. Buku lonse la ma parameter ndi malo oyeserera amoyo ali mu zolemba za API.

10Malire oona

Zinthu zinayi zoyenera kudziwa musanamange pa icho:

• Palibe masomphenya. Mawu okha; sichilandira zolowetsa za zithunzi pa endpoint iliyonse. Ngati mukufuna mtundu wosasidwa umene umaona, gwiritsani ntchito Kimi-K3, Kimi-K2.6, MiniMax-M3, MiMo-V2.5 kapena inkling.
• Malire apamwamba a ma token 16,384 a zotulutsa. Ma token a kulingalira amawerengedwa pa malirewo pa khama lapamwamba, choncho kulingalira kwautali kumasiya malo ochepa a yankho. Lumikizani mapempho pa ntchito zazikulu za kupanga.
• 3-bit ndi kufinya kwenikweni. Manambala oyambirira amafotokoza maziko, osati kapangidwe kameneka. Pa mlingo wapamwamba kwambiri wa kulondola, mitundu ya W8A16 (MiMo-V2.5-Pro, MiMo-V2.5, Hy3) imanyamula quantization yochepa kwambiri.
• Kosasidwa sikutanthauza kopanda zotsatirapo. Kusowa wosanjikiza wa kukana kumatanthauza kuti udindo wa chimene mukufunsa, ndi chimene mumachita ndi yankholo, ndi wanu — pansi pa mfundo zathu za ntchito ndi malamulo kumene mumagwira ntchito.

11Mafunso Ofunsidwa Kawirikawiri

---

Magwero: DeepSeek-AI — Kaadi ya mtundu ya DeepSeek-V4-Pro · Lipoti la ukadaulo la DeepSeek-V4 (CSA/HCA, mHC, Muon) · DeepSeek-V2 (MLA, DeepSeekMoE) · Lipoti la ukadaulo la DeepSeek-V3 · REAP the Experts (Cerebras Research) · Kufufuza kwa kusefa kwa Promptfoo · Discovering Forbidden Topics in Language Models. Manambala a mayeso ndi onenedwa ndi DeepSeek-AI a kutulutsidwa koyambirira; kutsimikizika kodziyimira pawokha kwa gulu lachitatu pa kapangidwe kofinyidwa kukudikirabe. Mitengo ndi mitengo ya mndandanda ya Shannon AI kuyambira 5 Seputembala 2026.