Ukonakaliswa kwe I-DeepSeek V4-Pro-Max ibingumcimbi we-seismic kwi-AI yokuvelisaKwiminyaka embalwa nje, siye sasuka ekulawuleni phantse ngokupheleleyo iimodeli zase-US, apho i-OpenAI, iGoogle, kunye ne-Anthropic zikhokela indlela, saya kwimeko apho inkampani yaseTshayina icebisa enye indlela evulelekileyo ekhuphisana ngqo neenkulu. I-United States yakhawuleza yafumana ithuba layo ngokuthintela ukufikelela kweTshayina kwiitships ze-AI eziphambili kunye noomatshini abafunekayo ukuze bazenze, nto leyo eqinisekisa ubunkokeli bayo phakathi kowama-2021 nowama-2024. Kodwa ekupheleni kowama-2024, uhlaselo lwaseTshayina lwaqala, kwaye ekuqaleni kowama-2025, kwavela umdlali omtsha: DeepSeek.
Ukususela kwinguqulelo yayo yokuqala, iDeepSeek sele ibenze abantu abakhulu abafana neNVIDIA baba nexhala, kuba ibonise ukuba kunokwenzeka ukufezekisa oku. Iziphumo ezithelekiswa ne-OpenAI zisebenzisa ixesha elincinci kakhulu lehardware kunye noqeqeshoUkuba "ulawulo" lweChatGPT lwaluchanekile ngenxa yamandla ekhompyutha angasemva kwayo, umyalezo wawucacile: iTshayina yayisenza into efanayo ngelixa ichitha imali encinci kakhulu. Kumvelisi weGPU ofana neNVIDIA, ukubona i-AI yanamhlanje iqeqeshwa ngamakhadi emizobo ambalwa kakhulu yayiyinto embi kakhulu. I-DeepSeek V4-Pro, kwaye ngaphezu kwako konke, uhlobo lwe-V4-Pro-MaxUbhejo lufikelele kumda: sithetha ngemodeli ye-AI evulekileyo ephucukileyo kwihlabathi kudidi lwayo, enoyilo oluthile kakhulu kunye nokugxila okuqatha kwingqiqo, iiarhente kunye nomxholo omde.
Imvelaphi ye-DeepSeek V4 Pro Max kunye nomxholo wokhuphiswano
I-DeepSeek ayizange ivele ndawoNgaphambi kwe-V4, besisele sibone izizukulwana ezininzi ezazisondela kancinci kancinci ekusebenzeni kweemodeli ezifana ne-GPT, i-Claude, kunye ne-Gemini. Ngowama-2025, uthotho lwe-DeepSeek 3.2 (kuquka iinguqulelo ezifana ne-Thinking kunye ne-Speciale) lugxile ekuphuculeni ubuchule bokuqiqa kunye ne-agent. Ezi nguqulelo bezisele zithelekiswa kakhulu ne-Anthropic, i-Google, kunye ne-OpenAI kwiibhentshi ezininzi, kodwa inkampani yaseTshayina yayinenyathelo elilandelayo elicacileyo: imodeli ye-V4 eyenzelwe ngokucacileyo ukuba iphumelele i-ChatGPT, uClaude, kunye noGemini kwimisebenzi ephambili.
Isiphumo saloo mzamo lusapho DeepSeek V4okungeyomodeli enye kodwa zimbini: I-V4-Pro y I-V4-FlashZombini zabelana ngefilosofi yoyilo, unikezelo lwe-token yesigidi esi-1, kunye nelayisensi ye-MIT evulekileyo, kodwa zineprofayili ezahlukeneyo kakhulu: I-V4-Pro ijolise ekusebenzeni okuphezulu kunye nokuqiqa okunzulu; I-V4-Flash ifuna ukusebenza kakuhle kunye nesantya esinomgangatho ofanayo ngokumangalisayo. Kule nkqubo yendalo, I-DeepSeek V4-Pro-Max Ivezwa njengolona hlobo luphezulu lokuqiqa lwe-V4-Pro, ityhala imodeli kwinqanaba layo eliphezulu kwimisebenzi enzima, yokhuphiswano, kunye neye-arhente.
Olu khuphiswano alusekho kwi-OpenAI okanye kwi-Google kuphela. Ngowama-2026, ezinye iimodeli zomda wemvelaphi evulekileyo nevaliweyo Abanye bajoyine olu gqatso, njengeKimi K 2.6 kunye neGLM-5.1, ezabelana ngeendawo eziphezulu kwiindlela ezininzi zokulinganisa. Nangona kunjalo, idatha ekhoyo ibonisa ukuba I-DeepSeek V4-Pro-Max iye yaba yeyona modeli ye-AI evelisa umthombo ovulekileyo ilungileyo kwiinkalo ezahlukeneyo, ingakumbi kwiinkqubo zokhuphiswano, ukuqiqa okuphucukileyo, kunye nemisebenzi yearhente ekumgangatho ophezulu.
Iinkcukacha zobugcisa zeDeepSeek V4-Pro kunye neV4-Pro-Max
Kwisiseko seDeepSeek V4-Pro sifumana Uyilo lweeNgcali eziDibeneyo (MoE) ludityaniswe neendlela zokhathalelo oluxutyiweyo kunye noyilo olukhethekileyo olubizwa ngokuba yiManifold-Constrained Hyper-Connections (mHC). Ingcinga icacile: ukukhulisa ukuya kwiitriliyoni zeeparamitha ngaphandle kokuba imodeli ingazinzi okanye ibize kakhulu ukubala, ngelixa igcina ifestile yomxholo esebenzisekayo yeethokheni ezili-1.000.000.
I-DeepSeek-V4-Pro ichazwa ngolu hlobo:
- I-Architecture: Umxube weeNgcali (MoE) + ingqwalasela exutyiweyo (Ingqwalasela eNcincisiweyo + Ingqwalasela eNcincisiweyo kakhulu) + uqhagamshelo lwe-mHC.
- Iiparameters zizonke: 1,6 trillion.
- Iiparameter ezisebenzayo ngethokheni nganye: 49.000 yeebhiliyoni zeerandi, nto leyo evumela ulungelelwaniso oluhle phakathi kwamandla kunye nokusebenza kakuhle.
- Ubude bomxholo: Iithokheni ezili-1.000.000 ngokuzenzekelayo, kunye neethokheni ezifikelela kwi-384.000 ezikhutshiweyo.
- Idatha yangaphambi kokuzilolonga: ngaphezulu kwee-token ezahlukeneyo ezingama-32 eebhiliyoni.
- IlayisenisiI-MIT, enezisindo ezivulekileyo nezizixhasayo ngokupheleleyo.
- Ukuchaneka kwamanani: umxube we-FP4 kwiingcali kunye ne-FP8 kwezinye izinto ezifunekayo, ubeka phambili inkumbulo kunye nokusebenza kakuhle kwekhompyutha.
- ubungakanani bokukhuphela: malunga ne-865 GB kwiisisindo ezipheleleyo.
- Qalisa: imboniso epapashwe nge-24 ka-Epreli, 2026.
Umahluko I-DeepSeek V4-Pro-Max Ayitshintshi inani lilonke leeparameter, kodwa ivumela indlela yokuqiqa ebukhali nenzulu kakhulu esekelwe kwi-V4-ProOlu luqwalaselo oluboniswa yinkampani ngokwayo njengencopho yesakhono sokuqiqa, ingakumbi xa ludityaniswe nemowudi yeThink Max. Kwiimilinganiselo zangaphakathi kunye nothelekiso kwiHugging Face, olu tshintsho lubekwe njengo imodeli evulekileyo enamandla kakhulu kwiimvavanyo ezininziphambi kwezinye iintsimbi ezivulekileyo kwaye zisondele kakhulu kwiimodeli zamva nje ezizimeleyo.
Le ndlela yenza i-DeepSeek V4-Pro kunye ne-V4-Pro-Max zibe zikhetho ezifanelekileyo kwi-intanethi. imisebenzi ebandakanya ucwangciso oluntsonkothileyo, izibalo eziphambili, iiarhente ezizimeleyo, kunye nohlalutyo olunzulu lwamaxwebhuapho ixabiso ngethokheni lithethelelwa kukuncipha kweempazamo kunye nomgangatho weempendulo.
I-DeepSeek V4-Flash: umntakwabo olula kodwa okwaziyo ukumangaza
Kunye neV4-Pro, iDeepSeek iqalise I-V4-Flashesingayichaza njenge "umntakwabo osebenzayo" wosapho. Ikwabelana ngesakhiwo esikumgangatho ophezulu, ilayisenisi, kunye nefestile yomxholo ofanayo, kodwa inciphisa ngamandla iiparameter ezipheleleyo nezisebenzayo ukuze ifumane isantya kwaye inciphise iindleko ngaphandle kokuwohloka okukhulu komgangatho.
Las Iinkcukacha eziphambili zeDeepSeek-V4-Flash Zizo:
- Iiparameters zizonke: 284.000 yezigidi.
- Iiparameter ezisebenzayo ngethokheni nganye: Iibhiliyoni ezili-13.000, phantse ikota yezo zeV4-Pro.
- Ubude bomxholo: Iithokheni eziyi-1.000.000, ezifana nePro.
- Ukuchaneka: indibaniselwano efanayo FP4 + FP8.
- ubungakanani bokukhuphela: malunga ne-160 GB, efikelelekayo ngakumbi kwiindlela zokusasaza zasekuhlaleni.
- Ixabiso kwi-API: I-$0,14 ngesigidi seethokheni ezifakiweyo kwaye i-$0,28 ngesigidi seethokheni eziphumileyo.
Into enomdla kukuba, nangona unayo phantse kathathu iiparameter zizonke zingaphantsi kuneDeepSeek 3.2I-V4-Flash iyayigqitha ngokuchanekileyo kwaye, kwiimvavanyo ezininzi, ingaphantsi nge-2-3% kuphela kune-V4-Pro. Nangona kunjalo, kwimisebenzi ethile umahluko uyabonakala: kwiimvavanyo ezifana ne IINKCUKACHA I-Parametric (EM) o I-SimpleQA-Verified (EM)Iziphumo zeFlash ezingaphaya kwe-doubles okanye eziphindwe kabini (umzekelo, ukusuka kwi-27,1% ukuya kwi-62,6% kwi-FACTS Parametric, okanye ukusuka kwi-28,3% ukuya kwi-55,2% kwi-SimpleQA-Verified). Esi sikhewu sibonisa ukuba I-Flash ayilomodeli ifanelekileyo xa ukuchaneka okugqithisileyo kweenyaniso kubalulekileKodwa ingena kakuhle kwizishwankathelo, incoko eqhelekileyo, kunye nokuveliswa kwekhowudi yemihla ngemihla.
Kwiziseko zophuhliso zangaphandle ezifana neClore.ai, ingcaciso icacile: I-V4-Flash yeyona ndawo imnandi kubasebenzisi abaninziKuba ingena kwi-1 × A100 80GB okanye 2 × RTX 4090 elinganisiweyo kwaye inika umlinganiselo olungileyo wexabiso/ukusebenza. I-DeepSeek V4-Pro, kwelinye icala, igcinelwe ukusetyenziswa okunzulu kakhulu ngezixhobo ezifana ne-8 × H100, 4 × H200 okanye 8 × B200 ene-NVLink, apho ingqalelo ye-hybrid inikwa khona kakhulu.
Uyilo lwe-hybrid kunye nokulungiswa kwangaphakathi: i-CSA, i-HCA, kunye ne-mHC
Inqanaba elikhulu lothotho lwe-V4 xa lithelekiswa ne-V3.2 likwi ulawulo lomxholo wexesha elide kunye nokuzinza koqeqeshoIiTransformers zakudala ziyahlupheka xa umxholo ukhula: ixabiso le-FLOPs kunye nememori ye-KV cache enyuka ngokukhawuleza. I-DeepSeek iphendula ngokudibanisa izinto ezintathu ezibalulekileyo: Ingqalelo eNcitshisiweyo (i-CSA), Ingqalelo Ecinezelwe Kakhulu (HCA) kunye Uqhagamshelo oluDibeneyo oluNxibelelanisiweyo (mHC).
La Ingqalelo Encinci Ecinezelweyo Isebenzisa ucinezelo olusekelwe kwi-token kwii-key-value pairs kwiimeko ezikude ngokuphakathi. Endaweni yokusebenzela yonke i-token yangaphambili, le modeli isebenza ngee-compressed and sparse representation, igcina ukuthembeka okufunekayo kodwa inciphisa kakhulu iimfuno zememori kunye nokubala. Ingqalelo Ecinezelwe KakhuluKwelinye icala, iya phambili ngakumbi: yenza ucinezelo olunamandla ngakumbi kwiithokheni ezikude kakhulu, igcina izishwankathelo ezimfutshane ezivumela imodeli ukuba "ikhumbule" into ebalulekileyo ngaphandle kokuthwala ubunzima obupheleleyo bembali.
Isiphumo esidibeneyo siyathandeka: Kwimeko yeethokheni ezi-1 yezigidi, iDeepSeek V4-Pro inciphisa ii-inference FLOPs ukuya kwi-27% yoko i-V3.2 ikufunayo. kwaye inciphisa imemori ye-KV cache ukuya kuthi ga kwi-10% xa kuthelekiswa nemodeli yangaphambili. Oku akupheleli nje kwingcamango; uya kuqaphela umahluko xa uzalisa amaxwebhu amakhulu okanye ucubungula yonke ikhowudi yokugcina ngaphandle kokuba iGPU ingasebenzi kakuhle.
Las Uqhagamshelo oluDibeneyo oluNxibelelanisiweyo (mHC) Zithatha indawo yoqhagamshelo oluqhelekileyo oluqhelekileyo lweetransformer. Ngokukhawulela uhlaziyo lobunzima kwiRiemannian manifold, i-mHC Iphucula ukusasazwa kwesignali ngamakhulu eeleya. kwaye ivumela imodeli eneeparameter eziyi-1,6 trillion ukuba iqeqeshwe ngokuzinzileyo. Injongo kukuqinisekisa ukuba ubunzulu obugqithisileyo bemodeli abukhokeleli kwiingxaki zokuqhuma okanye zokuphela kwe-gradient.
Ekugqibeleni, iDeepSeek yazisa I-Muon optimizer (Momentum + Orthogonalization) endaweni ka-AdamW. I-Muon i-orthogonalize uhlaziyo lwe-gradient phakathi kwamanyathelo alandelelanayo, alandelayo Iyasusa ukuphinda-phinda, ikhawulezise ukuhlangana, kwaye inika uzinzo. xa usebenza ngee-token ezingaphezulu kweebhiliyoni ezingama-32 zoqeqesho lwangaphambi koqeqesho. Yinto ebalulekileyo ekulinganiseni isangqa "seeparameter ezininzi, umxholo omninzi, kunye noqeqesho oluyinyani ngokwexesha kunye nezixhobo."
Iindlela zokuqiqa: Ukungacingi, Ukucinga Ngezinto Eziphezulu, kunye nokuCinga Ngezinto Eziphezulu
Enye yezinto ezahlula i-DeepSeek V4-Pro kunye ne-V4-Pro-Max kukuba ziyabandakanya iindlela ezintathu zokuqiqa ezinokulungiselelwa nge-APIEzi zenzelwe ukulungisa umzamo "wochungechunge lweengcinga" ngokwemisebenzi. Ayizizo zonke imibuzo ezifuna ingqiqo enzulu, kwaye ukuba uhlala uyivula, uya kuyinyusa ngokungeyomfuneko i-token bill yakho.
Iindlela ezintathu zezi:
- Ukungacingi: impendulo ethe ngqo, ngaphandle kwenkqubo yokucinga ecacileyo. Kwi-API, ilungiselelwe nge ukucinga: {uhlobo: “okhubazekileyo”}Yindlela eqhelekileyo yencoko elula, izishwankathelo ezikhawulezayo, okanye ukuvelisa umbhalo apho isantya sibaluleke kakhulu.
- Cinga Ngezinto Eziphezulu: ingqiqo ecwangcisiweyo kunye uhlahlo lwabiwo-mali oluchaziweyo lwethokheniUmzekelo, ukucinga: {type: «enabled», budget_token: N}. Oku kusetyenziswa xa sifuna ingcaciso eqinileyo ngaphandle kokuya kwiindlela ezigqithisileyo.
- Cinga ngoMax: ingqiqo epheleleyo neyandisiweyo, enxulumene ne isindululo senkqubo ekhethekileyo kunye noqwalaselo lokucinga: {uhlobo: «max»}. Yenzelwe iimeko ezinomxholo obanzi kakhulu (iithokheni ezisebenzayo ezingama-384K+) kunye nemisebenzi enzima kakhulu.
Iimilinganiselo ezisemthethweni zibonisa umahluko omkhulu phakathi kwe-Non-think kunye ne-Think Max. Umzekelo, kwi- LiveCodeBench I-V4-Pro isuka kwi-56,8% ukuya kwi-93,5%, kwi Idayimani yeGPQA ukusuka kwi-72,9% ukuya kwi-90,1%, nakwi- I-HMMT 2026 ngoFebruwari Itsiba ukusuka kwi-31,7% ukuya kwi-95,2%. Kwinkqubo yokhuphiswano, Uvavanyo lweCodeforces lufikelela kumanqaku angama-3206 kwimo yokuqiqa ephezulu, embeka phakathi kwabathathi-nxaxheba ababalaseleyo kwaye imenza, ngokwedatha yephepha, imodeli yokuqala evulekileyo ekwaziyo ukufanisa i-GPT-5.4 kuloo msebenzi uthile.
Olu yilo lwenza iDeepSeek V4-Pro-Max ibe nomtsalane ngakumbi kuyo iiarhente zeenkqubo, uyilo lwe-algorithm oluntsonkothileyo, ukusombulula iingxaki ze-STEM eziphambili, kunye novavanyo lweshishiniapho sinokutshintshela kwimo yeThink Max kuphela xa umsebenzi usifuna, sigcine iNon-think okanye iThink High yonke enye into.
Ukusebenza kwiindlela zokulinganisa, ukuqiqa, kunye neearhente
Ngokuphathelele ubuchule obucocekileyo, iDeepSeek V4-Pro kunye, ngokongeza, iV4-Pro-Max, zigqwesile ngakumbi kwiinkqubo, ekuqiqeni okuphezulu, nakwimisebenzi enzima ye-arhente. Kwinkqubo, imodeli iyabonakala. 93,5% kwiLiveCodeBench (Pass@1), Isisombululo se-80,6% kwi-SWE-bench Iqinisekisiwe, I-55,4% kwi-SWE-bench Pro y 76,2% kwi-SWE-benchNgaphezu koko, idibana ngokwendalo neenkqubo ezifana Ikhowudi kaClaude, i-OpenClaw kunye ne-OpenCode, iqinisa ukugxila kwayo kwiiarhente ezibhala, eziqhuba, nezilungisa ikhowudi.
Kwingqiqo nolwazi, amanani ahamba kunye: I-MMLU-Pro 87,5% kwi-Think Max, Idayimani yeGPQA 90,1%, I-HLE 37,7% y I-SimpleQA-Iqinisekisiwe yi-57,9%. Kwimeko ye MMMLU (iilwimi ezininzi)Isiseko sale modeli sifikelela malunga ne-90,3%, nto leyo ethetha ukuba Ukwazi kakuhle iilwimi ezininzi, kuquka iSpanishUkudibana kwezi ziphumo kubeka i-V4-Pro phezulu kuluhlu lweemodeli ezivulelekileyo, kufutshane kakhulu neemodeli ezivulelekileyo eziphezulu.
Kwimeko yexesha elide, iDeepSeek V4 ikhanya ngakumbi: I-83,5% kwi-MRCR 1M (inaliti ikwisixa seengca)apho isebenza ngcono kuneGemini 3.1-Pro, kwaye I-62,0% kwiCorpusQA 1M kwimo yeThink Maxngelinye lawona manqaku abalaseleyo ngaphandle kwenkqubo yendalo iClaude. I-LongBench-V2Isiseko sijikeleza malunga ne-51,5%, nto leyo eqinisa ingcamango yokuba imodeli yenzelwe ukufunda, ukugcina, nokuqiqa ngemiqulu emikhulu yombhalo.
Las imisebenzi yobumeli Zezinye zezona zinto zinamandla: 67,9% kwi Ibhentshi yesiphelo sendlela 2.0 (Cinga nge-Max mode), 80,6% kwi-SWE-bench Iqinisekisiwe, 73,6% kwi I-MCPAtlas yoluntu, 83,4% kwi Khangela kwiComp kunye ne-51,8% kwi I-Toothlonrhoqo kwiindlela eziphezulu zokuqiqa. Olu lwazi luxhasa ibango lokuba i-V4-Pro-Max ngumgqatswa obalaseleyo we- iiarhente ezizimeleyo ezidibanisa izixhobo ezininzi, imiyalelo yesiphelo, kunye neefowuni ze-API kwimisebenzi yezinyathelo ezininzi.
Uthelekiso lwe-V4-Pro vs V4-Flash kunye necebo lokusebenzisa
I-DeepSeek ayithathi isigqibo ngokuthi "sebenzisa i-V4 okanye hayi," kodwa njenge Ukukhetha i-V4-Pro kunye ne-V4-Flash efanelekileyo kuhlobo ngalunye lomsebenziEnyanisweni, zombini ezi modeli zine-API efanayo, zihambelana nefomathi ye-OpenAI's Chat Completions kunye ne-Anthropic's messaging protocol, kunye neendlela zazo zokuqiqa. Ukutshintsha phakathi kwazo kudla ngokuba lula njengokutshintsha i-ID yomzekelo kwifowuni.
Le nkampani ngokwayo icwangcisa imisebenzi yayo ngolu hlobo:
- I-V4-Pro: ulwazi olubanzi lwehlabathi, ukuqiqa okukumgangatho wehlabathi kwizibalo, i-STEM kunye nenkqubo, imodeli enamandla kwimisebenzi enzima yearhente.
- I-V4-FlashUkuqiqa okufana kakhulu noko kwePro: ukusebenza okufanayo kwimisebenzi elula yearhente, ukusebenza okuphantsi kwimisebenzi enzima. Ixabiso liphantsi ukuyikhonza kwaye ineempendulo ezikhawulezayo.
Kwezinye iimvavanyo umsantsa uncinci (amanqaku ali-1-3) kwaye kwezinye uyanda kakhulu. Umzekelo, kwi I-MMLU-Pro, i-LiveCodeBench okanye i-SWE-Verified umahluko mncinci, ngelixa ku I-SimpleQA-Iqinisekisiwe o Ibhentshi yesiphelo sendlela 2.0 Umgama uba ngamanani amabini. Ufundo olusebenzayo lucacile: I-Flash ilungele imisebenzi elula nephakathi (incoko, isishwankathelo, ukuhlelwa, ukuveliswa kwekhowudi ngqo), kodwa Akusebenzi kakuhle xa sithetha ngezinto ezintsonkothileyo, ukufunyanwa kweenyani ezichanekileyo kakhulu, okanye iimeko ezinobungozi kwezoshishino..
Icebo elisengqiqweni kakhulu leemveliso ze-AI kukuba indlela ye-hybridThumela zonke izicelo kwi-V4-Flash ngokuzenzekelayo uze unyuselwe kwi-V4-Pro (okanye kwi-V4-Pro-Max kwi-ThinkMax) kuphela xa umsebenzi ubangela iimeko ezithile: ukungaphumeleli kwefowuni yesixhobo, umda wokuthembana ongafezekiswanga, impendulo yokuphawula yomsebenzisi njengengalunganga, okanye isigaba somsebenzi esichongiweyo njengesibalulekileyo. Ukuba ipesenti yezicelo eziya kwi-Pro zihlala ziphantsi (umzekelo, ngaphantsi kwe-5-10%), i-Flash iya kuba ngumsebenzi onzima kwaye i-Pro ibe "yikhadi lasendle" kwiimeko ezinzima.
Iindleko ze-API, ukuthunyelwa, kunye nokuhambelana
Kwinqanaba elisemthethweni le-DeepSeek API, Amaxabiso abonisa ngokucacileyo indawo ebekwe kuyo imodeli nganye.Kwi-V4-Flash, umrhumo wokungena umalunga ne-$0,14 ngesigidi seethokheni ($0,028 ukuba kukho i-cache hit kwisiqalo), kwaye umrhumo wokuphuma yi-$0,28 ngesigidi seethokheni. Kwi-V4-Pro, la manani anyuka aye kwi-$1,74 ngesigidi seethokheni zokungena ($0,145 nge-cache) kunye ne-$3,48 ngesigidi seethokheni zokuphuma. Ulwakhiwo lwamaxabiso luyafana; yintoni utshintsho oluyintlawulo: I-Pro ibiza ngaphezulu ngokuphindwe kalishumi ukuya kwi-12 ithokheni nganye eyenziweyo.
Zombini iimodeli zibonakaliswa nge-API efanayo (api.deepseek.com/v1kwaye wabelane ngokuhambelana Ukugqitywa kweNgxoxo ye-OpenAI, ifomathi yemiyalezo ye-Anthropic, ukusasazwa kweempendulo, kunye nomxholo wokuqiqa kwiindlela zeThink High kunye neThink Max. Ukuhambisa indlela phakathi kweemodeli ezimbini kuyinto engenamsebenzi, okwenza kube lula kakhulu ukuzama iindlela ezahlukeneyo zokudibanisa.
Ngaphaya kwe-API esemthethweni, i-DeepSeek V4 iyakwazi Khuphela kwiHugging Face kunye neModelScope for ukusasazwa kwendawoOku kuvula ithuba lokuyisebenzisa kwisiseko sakho okanye kubaboneleli beenkcukacha-manani beqela lesithathu, abanjengoNovita, Clore.ai, okanye iAtlasCloud, esele bebonelela ngeziganeko ezilungiselelwe kwangaphambili. Kwi-V4-Flash, i-Unsloth ipapashe ubuninzi be-GGUF obuvumela imodeli ukuba isebenze kwizixhobo ezingabizi kakhulu (umzekelo, i-80 GB GPU enye okanye ii-48 GB GPU ezimbini) ezinomgangatho ophantse ube yi-FP8 usebenzisa i-Q4_K_M.
Ngokuphathelele izikhokelo, I-vLLM 0.7.x inika inkxaso yosuku olungenazo zombini iindawo zokujongaNgee-kernel ze-hybrid attention ezifuna ukhetho lwe-`--trust-remote-code` kunye ne-hardware ye-Hopper okanye ye-Blackwell ukuze kufezekiswe isantya esiphezulu, i-SGLang yenye indlela enomdla. I-RadixAttention yayo kunye ne-prefix caching zisebenza kakuhle nge-hybrid attention kwaye zihlala zibonelela ngokusebenza okungcono kwi-Hopper GPUs, ngakumbi kwimisebenzi ye-agent ene-shared prompts.
Amatyala okusetyenziswa acetyiswayo kunye nokufuduka okuvela kwiimodeli zangaphambili
Amaxwebhu avela kwi-DeepSeek kunye namaqonga aliqela adibanisa i-V4 ngokwayo acebisa isikhokelo esibonisa ukuba loluphi uhlobo kunye nendlela yokuqiqa omawuyisebenzise kuhlobo ngalunye lomsebenzi. Nje:
- Ingxoxo kunye nemibuzo ngokubanzi: V4-Flash kwimo yokungacingi, isantya kunye nexabiso eliphantsi zibekwe phambili.
- Ukugqitywa kwekhowudi eqhelekileyo: V4-Flash Ayicingi, apho ukubambezeleka kubalulekile.
- Uyilo lwee-algorithms ezintsonkothileyo: V4-Pro eneThink High, ifuna ulungelelwano phakathi kokuchaneka kunye nexesha.
- Inkqubo yokhuphiswano: I-V4-Pro-Max kwi-Think Max, ukuze isebenzise ngokupheleleyo amandla ayo.
- Isishwankathelo soxwebhu olukhulu: V4-Flash Ayicingi nokucinga, ilungele ukulayisha umthamo.
- Uhlalutyo olunzulu lwamaxwebhu olunemixholo emininzi: V4-Pro Think High, esebenzisa izigidi zeethokheni kunye nokuqiqa okucwangcisiweyo ngakumbi.
- Iiarhente ezizimeleyo ezintsonkothileyo: V4-Pro-Max Think Max, kwimisebenzi yokusebenza enamanyathelo amaninzi kunye neyomngcipheko ophezulu.
Amaqonga emveli afana I-Framia.pro Sele besebenzisa uthungelwano olukrelekrele phakathi kwezi zicwangciso, belungisa ngokuzenzekelayo imo yotshintsho kunye ne-V4 ngokusekelwe kubunzima bomsebenzi, ukuze baphucule umgangatho, iindleko kunye namaxesha okuphendula ngaphakathi kwemisebenzi yokudala nophuhliso.
Ukuba uvela kwiimodeli zangaphambili ezifana I-DeepSeek V3 okanye i-DeepSeek-R1Ukufuduka kulula kakhulu: imodeli efanayo yosapho kunye netemplate yencoko ziyagcinwa, kwaye ungenza utshintsho oluthe ngqo kwi-vLLM okanye kwezinye iiframeworks ezihambelanayo. Ngaphezu koko, ii-ID ezindala deepseek-incoko y deepseek-reasoner Sele zithunyelwa kwi-V4-Flash, kunye neshedyuli yomhlalaphantsi emiselwe umhla wama-24 kuJulayi 2026. Oko kuthetha ukuba abasebenzisi abaninzi sele benandipha umgangatho we-V4-Flash ngaphandle kokuba neemodeli ezitshintshiweyo ngokucacileyo.
Ngokwembono yeshishini, kubalulekile ukuqwalasela umxholo wezopolitiko kunye nolawuloEzinye iindlela zokusasazwa ngaphandle kweTshayina zinokuvavanywa ngakumbi ukuba kusetyenziswa i-API esemthethweni. Kwiimeko ezinjalo, ukuzibamba ngokwakho ii-open weights ezinelayisensi ye-MIT kuba lolona khetho lucocekileyo lokuthobela imigaqo nemigaqo-nkqubo yangaphakathi.
Ngenxa yazo zonke ezi zinto zingentla, iDeepSeek V4-Pro-Max izibeka njenge imodeli ye-open weight frontier ekwaziyo kakhulu ekhoyo ngokuUkudibanisa uyilo olucwangcisiweyo kakuhle lweemeko zexesha elide, iindlela zokuqiqa eziguquguqukayo, ukusebenza okuphezulu kwiinkqubo nakwiiarhente, kunye namaxabiso, kwiimeko ezininzi, akhuphisana ngokuphindwe kalishumi ukuya kuma-35 kunezinye iindlela zobunini, i-V4-Pro-Max kunye nenkqubo yayo yendalo imele inyathelo elibalulekileyo kuye nabani na ofuna i-AI ekrelekrele ngokwenene ekwaziyo ukuqiqa ngokunzulu, ukucubungula iiprojekthi ezinkulu, kunye nokusebenza njengonjineli ophezulu wedijithali.
