Isizukulwane esisha samamodeli se-DeepSeek sesibe yindawo ebalulekile engxoxweni yezobuchwepheshe ngesiphakamiso esicacile kakhulu: umongo wamathokheni afinyelela esigidini kanye nokwakheka kwamapharamitha angaphezu kwe-trillion eyodwa Yakhelwe ukusebenza kahle futhi, ngaphezu kwakho konke, ishibhile kakhulu kunezinye izindlela ezivaliwe e-United States, inkampani yaseShayina isebenzise ngokuphelele i-V4, umndeni ohlanganisa izinsimbi ezivulekile, iwindi elikhulu lokuqukethwe, kanye nesu lokubeka amanani ngendlela ebukhali.
Lesi sinyathelo sifika ngesikhathi lapho iYurophu neSpain zihlola izindleko kanye nobukhosi bezobuchwepheshe be-AI. I-DeepSeek V4 iziveza njengenketho ekhangayo yezinkampani ezintsha zaseYurophu, ama-SME kanye nezinkampani ezinkulu ezidinga amakhono ezingeni lomngcele, kodwa azikwazi—noma azifuni—ukuthembela ngokuphelele kuma-API abanikazi abizayo noma ihadiwe ekhethekile njengama-GPU e-NVIDIA afunwa kakhulu.
Umndeni we-V4 ogxile ku-1T yamapharamitha kanye nomongo wamathokheni angu-1M
I-DeepSeek imemezele ukufika kwe-DeepSeek-V4 Preview njengomndeni wamamodeli avulekile agxile emibonweni emibili: ifasitela lomongo lamathokheni afinyelela ku-1 million kanye nezakhiwo ezinkulu ezisekelwe ku-Mixture-of-Experts (MoE)Kulo mndeni, izinhlobo ezimbili eziyinhloko zivelele: i-DeepSeek-V4-Pro kanye ne-DeepSeek-V4-Flash, zombili ezinalowo mongo we-1M njengophawu.
Ekugcineni okunesifiso esikhulu, i-V4-Pro isebenza ngezibalo ze- kufika ku-1,6 trillion iyonke yamapharamitha (1,6T), yize isebenza kuphela phakathi kwamapharamitha ayizigidigidi ezingama-32 nezingama-49 esinyathelweni ngasinye sokunquma ngenxa yohlelo lwe-MoE, olubalulekile ekugcineni ukusebenza kahle. Ngesikhathi esifanayo, inkampani yethule izinhlobo ezilula, njenge-V4-Flash kanye ne-V4-Lite, ezinamapharamitha aphelele angaba yizigidigidi ezingama-284-285 kanye namapharamitha asebenzayo angaba yizigidigidi eziyi-13, aklanyelwe ukuthunyelwa lapho isivinini kanye nezindleko kuyizinto eziza kuqala.
Inani eliphelele lamapharamitha libeka umndeni we-V4 phezulu emakethe, kodwa imininingwane ebalulekile ukuthi Ingxenye encane kuphela yalabo ochwepheshe esetshenziswa yi-token-activated.Lokhu kuyivumela ukuthi iziphathe njengemodeli enkulu ngokwezinga lomthamo, kodwa ngokusetshenziswa kwamandla okubala okuseduze nokwamamodeli amancane kakhulu. Kuyindlela ehambisana nendaba kaDeepSeek: ukuncintisana namamodeli amakhulu, anomthombo ovaliwe ngaphandle kokuphakamisa izindleko zokusebenzisa.
Le nkampani iphinde yakhipha izinhlobo zokuqala ezifana ne-V4-Lite, ezisebenza njengokuqinisekiswa kobuchwepheshe, futhi ibilokhu ilungisa ishejuli yokuthunyelwa. I-V4 isesesigabeni sokuhlola esilinganiselwe Kwezinye izimo, umndeni we-V4 Preview ungasetshenziswa kakade ku-chatbot esemthethweni kanye nange-API ebuyekeziwe yenkampani, kanye nomongo we-1M njengenani elizenzakalelayo kumasevisi ayo.
Ukwakhiwa kwezakhiwo okuhlanganisiwe kanye nenhlanganisela yochwepheshe ukuze kwenziwe umongo wesikhathi eside usebenze kahle
Isihluthulelo sekhono le-DeepSeek lokunikeza iwindi lokuqukethwe lamathokheni ayisigidi ngaphandle kokunyuka kwezindleko zokubikezela sisekwakhekeni kwayo. Umenzi uchaza ukuthi i-V4 iyethula inhlanganisela yokunakekelwa okuxubile, inhlanganisela yochwepheshe kanye namasu okucindezela yakhelwe ukusebenza ngezilandelaniso ezinde kakhulu, inciphisa ama-FLOP ngethokheni kanye nememori edingekayo.
Phakathi kwezingxenye zobuchwepheshe ezikhulunywa yinkampani, okulandelayo kuyagqama: I-MLA (Ukunakekelwa Okucashile Okunezinhloko Eziningi), i-DSA noma i-DeepSeek Sparse Attention, kanye nezinqubo zememori ezinemibandela ezifana ne-EngramNdawonye, lezi zingxenye zihlose ukunciphisa umthwalo wokubala ukunaka, ikakhulukazi lapho imodeli kufanele iphathe amakhulu ezinkulungwane noma amathokheni ayisigidi ngesikhathi esisodwa.
Ngokusho kwedatha eyabiwe yinkampani uqobo, ezimweni zamathokheni angu-1M I-DeepSeek-V4-Pro ingadinga cishe ama-27% ama-FLOP ngethokheni kanye no-10% kuphela we-KV cache uma kuqhathaniswa nezinguqulo zangaphambilini ezifana ne-DeepSeek-V3.2Izinhlobo ezilula, njenge-V4-Flash, ziyanciphisa lezi zibalo, zizibeka njengezixazululo zokuphetha okusheshayo kwezinhlelo zokusebenza lapho ukubambezeleka kubalulekile khona.
Lezi zinhlobo zokuthuthuka azigcini nje ngokwemfundiso: inkampani ithi inhlanganisela ye-MoE, ukunaka okuhlakazekile, kanye nokuqonda umongo kuvumela isebenza ngomongo omude kakhulu ku ihadiwe engeqisi kakhulu kakade izindleko ngamathokheni ayisigidi ziphansi kakhulu kunezamamodeli amaningi avaliwe anama-windows angamathokheni angu-128K noma angu-200K.
Ukusebenza kahle emisebenzini yokucabanga, ukuhlela, kanye neyokwenza izinto
I-DeepSeek ayifuni nje ukuvelela ngenxa yobukhulu bayo kanye nomongo wayo. Ekuqhathaniseni kwayo kwangaphakathi, inkampani igcizelela ukuthi I-V4-Pro kanye nezinhlobo zayo zilungiselelwe ngokukhethekile ukucabanga okuyinkimbinkimbi, izinhlelo, kanye nama-ejentiLezi zindawo ezintathu okwamanje zibangela ingxenye enkulu yesidingo sebhizinisi. Izilinganiso ezifana ne-SWE-bench, ezenzelwe ukukala amandla Ukuqonda nokushintsha ama-repository ekhodiKukhona inkulumo ngezibalo ezingaphezu kuka-80% ngokunemba, ngokuhambisana namamodeli avaliwe ahamba phambili.
Ekucabangeni okuvamile—kufaka phakathi izibalo, izifundo ze-STEM, kanye nezinkinga zochungechunge lwemicabango—inkampani ibeka i-V4-Pro njengenye yamamodeli avulekile aqine kakhulufuthi ithi isondela ezingeni leziphakamiso zemingcele evaliwe. Ngokuphathelene nokuqwashisa umhlaba wonke, idatha yangaphakathi iyibeka phambili ohlelweni oluvulekile lwe-ecosystem futhi ngemuva kwamamodeli ambalwa akhethekile, njenge izinhlobo ezithile ezithuthukisiwe ze-Gemini.
Ngale kwezinombolo, kugcizelelwa kakhulu imisebenzi ye-ejensi Kukhomba ukusetshenziswa okudlulela ngale kwengxoxo eyisisekelo. I-DeepSeek ithi lokho I-V4 isivele iqhuba ingqalasizinda yayo yama-ejenti ekhodi kanye nezinhlelo ezilandelana ngezinyathelo eziningiBafinyelela amathuluzi futhi basebenza ezindaweni zokugcina izinto ezinkulu noma ezigciniwe zemibhalo. Le ndlela ihambisana nomkhuba wamanje embonini, lapho izinkampani eziningi zingasafuni nje i-chatbot, kodwa zifuna abasizi abakwazi ukusebenza “njengozakwabo bedijithali” ngaphakathi kwemisebenzi eyinkimbinkimbi.
Lokhu kuqhathanisa kufanele kuthathwe ngokuhlakanipha: njengakuyo yonke imibiko ye-AI yakamuva, Iningi ledatha livela enkampanini uqobo kanye nasezivivinyweni ezilawulwayo.Noma kunjalo, inhlanganisela yomongo omude, ukwakheka okusebenzayo, kanye nokusebenza okuncintisanayo kudala ukunaka phakathi kwabathuthukisi baseYurophu abaqhathanisa izindleko namakhono nezinketho ezifana ne-GPT, i-Claude, i-Llama, noma i-Mistral.
Amamodeli avulekile, izisindo ezishicilelwe, kanye nokuhambisana nama-API adumile
Esinye sezici ezibalulekile ezilethe i-DeepSeek udumo lwayo ukuzibophezela kwayo ohlelweni oluvulekile. Nge-V4, inkampani iqinisa le ndlela: ushicilele umbiko wezobuchwepheshe futhi wakhipha izinkulumo ezivulekile zomndeni ezinkundleni ezifana ne-Hugging Face.okuvumela abacwaningi, izinkampani kanye nohulumeni ukuthi balande amamodeli futhi bawasebenzise kwingqalasizinda yabo.
Le ndlela yokuvula iminyango, ngokungafani neziphakamiso ezivaliwe ngokuphelele zamalebhu amaningi ase-US, inemiphumela ecacile ngeSpain kanye ne-European Union. Amathuba okusebenzisa la mamodeli ku izikhungo zedatha ngaphakathi kwe-EUngaphansi kwezinhlaka ezifana ne-GDPR kanye nomthethonqubo we-AI wesikhathi esizayo we-EUInikeza indlela yokugcina ukulawula okukhulu kwedatha ngaphandle kokulahla amakhono aphezulu.
Ngokuphathelene nokuhlanganiswa okusebenzayo, i-DeepSeek ikhethe ukunciphisa ukungqubuzana: I-API igcina i-base_url efanayo futhi iyahambisana nezinhlelo ze-OpenAI's ChatCompletions kanye ne- Izixhumi ze-AnthropicKumaqembu amaningi okuthuthukisa, lokhu kusho ukuthi ukuthuthela izivivinyo noma izingxenye zethrafikhi ku-V4 kunqunyelwe ekushintsheni isihlonzi semodeli sibe yi-deepseek-v4-pro noma i-deepseek-v4-flash nokulungisa amapharamitha ambalwa.
Ngesikhathi esifanayo, inkampani ibeke isikhathi sokuthatha umhlalaphansi kumamodeli amadala, njenge-deepseek-chat kanye ne-deepseek-reasoner. Zizomiswa futhi ziqondiswe ku-V4-Flash kuze kube yilapho behoxiswa ngokuphelele, okuphoqa labo abasebenzise bona ukuthi baqale ukulungiselela ukufuduka. Kuyindlela ecacile yokugxila ekunikezeni isizukulwane esisha nokugwema ukuhlukanisa isisekelo sabasebenzisi sibe izinhlobo eziningi kakhulu zefa.
Izindleko zokuqagela eziqukethwe kanye nokugxila ekusebenzeni kahle kwezomnotho
Indaba ka-DeepSeek ibilokhu igxile ekusebenzeni kahle selokhu yasungulwa. Nge-V4, leyo nkulumo iqiniswa yinhlanganisela yokwakheka kwe-MoE, ukunakwa okusabalalisiwe, kanye nokwenza ngcono ihadiwe okuhlose... yehlisa izindleko ngamathokheni ayisigidi ziye emazingeni angaphansi kakhulu kwalawo ama-API e-premium aziwa kakhuluEzinye izihlaziyi zangaphandle zikhuluma ngezibalo ezicishe zibe ngu-$0,30 ngezigidi zamathokheni okungena kwezinye izilungiselelo, ingxenye encane yalokho amamodeli avaliwe aphezulu akhokhisa.
Esimweni saseYurophu, lapho izindleko zengqalasizinda kanye namandla zibalulekile khona, lokhu kugxila ekusebenzeni kahle kuhambisana kahle nezidingo zamabhizinisi asafufusa kanye nama-SME. Ukucubungula amadokhumenti asemthethweni abanzi, amarekhodi ezokwelapha amade, noma yonke indawo yokugcina isofthiwe Akusekho ukuba yindawo yokunethezeka egcinelwe izinkampani ezinesabelomali esingenamkhawulo futhi iba yingxenye yezimo ezingabizi kakhulu zamaphrojekthi asafufusa.
Abanye abahlinzeki bengqalasizinda ye-AI sebevele banikeza ukufinyelela kwangaphambi kwesikhathi kuma-node asekelwe ku-DeepSeek V4 njengengxenye yamakhathalogi abo, okwenza kube lula ezinkampanini zaseYurophu bangahlola ukusebenza kwangempela kanye nezindleko ngaphandle kokudinga ukwakha ingqalasizinda yabo kusukela ekuqaleni.Ezinhlanganweni eziningi, lesi sigaba sokuhlola siyisinyathelo sokuqala ngaphambi kokunquma ukuthi uzoqhubeka yini nomodeli othunyelwe ngaphandle noma ukhethe ukuthunyelwa endaweni.
Okwamanje, ukuthula kwenkampani okuncane mayelana nezindleko zokuqeqesha eziqondile kanye nehadiwe ethile esetshenzisiwe kuphakamise ukungabaza kwezinye imikhakha. Kusukela ngo-2025, izinsolo ziye zasakazeka mayelana nenani langempela lezinsiza ezidingekayo ukuqeqesha amamodeli ayo, okuhlanganisa nezilinganiso ezikhomba amashumi ezinkulungwane zama-GPU aphezulu. I-DeepSeek igcizelela ukuthi isifinyelele isigaba esisha "somongo wesikhathi eside onenzuzo"Kodwa ayikakakucacisi ngokuphelele okungaziwa mayelana nobukhulu bezinto ezisetshenziswayo.
Umthelela kuma-startups kanye nezinkampani eSpain naseYurophu
Kuhlelo losomabhizinisi lwaseYurophu, ikakhulukazi izinkampani ezintsha zobuchwepheshe eSpain, ukuvela kwamamodeli afana ne-DeepSeek V4 kuvula izinketho okwakunzima ukuzicabangela kuze kube muva nje. Finyelela imodeli enamapharamitha angaphezu kwe-trillion ngaphakathi komongo wamathokheni angu-1M kanye nezisindo ezivulekile Ikuvumela ukuthi uhlole imikhiqizo ethuthukisiwe ngaphandle kokuthembela kuphela kubaphakeli beSilicon Valley.
Emikhakheni elawulwayo—ezezimali, ezempilo, ezomthetho, ezokuphathwa komphakathi—amathuba okuba sebenzisa imodeli ezikhungweni zedatha ngaphakathi kwe-EU noma ngisho nasezikhungweni zakho Lokhu kubaluleke kakhulu. Ukuthobela imithetho ye-GDPR kanye nemithethonqubo kazwelonke yokuvikelwa kwedatha kuba lula ukuyilawula lapho ulwazi kungadingeki lushiye izindawo zaseYurophu ukuze lucutshungulwe ngemodeli ye-AI.
Izinkampani ezintsha zaseSpain ezisebenza ngemibhalo eminingi, njenge-legaltech, i-healthtech, noma amathuluzi onjiniyela, zingasebenzisa umongo wamathokheni angu-1M ukuze ukuhlaziya amafayela aphelele, umlando wezokwelapha omude kakhulu, noma izindawo zokugcina amakhodi ezi-monolithic ngaphandle kokudinga ukuzihlukanisa zibe yizicucu eziningi futhi uklame izinhlelo zokubuyisa eziyinkimbinkimbi. Lokhu kunciphisa ubunzima bobuchwepheshe, futhi ezimweni eziningi, nokubambezeleka.
Ngesikhathi esifanayo, kubalulekile ukukhumbula izingozi: uhlelo lwamathuluzi azungeze i-DeepSeek luncane kunolwamanye amamodeli avulekile njenge-Llama, futhi Imibhalo kanye nokusekelwa komphakathi kusakhulaNgaphezu kwalokho, iqiniso lokuthi iyinkampani yaseShayina lethula ingxenye yezepolitiki yendawo ezinye izinhlangano zaseYurophu eziyibheka ngokuqapha, ikakhulukazi kumaphrojekthi ahlobene nokuphathwa noma ingqalasizinda ebalulekile.
Isinyathelo esibeka ingcindezi kumamodeli abiza kakhulu navaliwe
Ngaphandle kwemininingwane yayo ethile, i-DeepSeek V4 ihunyushwa ngaphakathi komkhakha ngokuthi isinyathelo esengeziwe ekucindezelweni kokuncintisana kumamodeli avaliwe abizayo kakhulu emaketheNgokusungula umongo wethokheni ye-1M njengojwayelekile kuzo zonke izinsizakalo zayo ezisemthethweni futhi uhambisane nawo nezisindo ezivulekile, inkampani yaseShayina ithumela umyalezo ocacile: umongo omude kakhulu awusadingeki ube yisici esikhethekile samamodeli ambalwa akhethekile anentengo ephezulu.
Kuma-laboratory amakhulu aseNtshonalanga, lokhu kuletha inselele. I-OpenAI, i-Anthropic, kanye ne-Google bezilokhu zisebenzisa inhlanganisela ye- ikhwalithi ephezulu, umongo obanzi kanye nesimiso sezinto eziphilayo esinobungcweti njengesiphakamiso senani. Ukuvela kwenketho evulekile enomongo ophakeme nakakhulu kwezinye izimo kanye nezindleko eziphansi kakhulu kuphoqa ukuthi kucatshangelwe kabusha amasu omkhiqizo namanani, ikakhulukazi ezigabeni lapho umkhawulo wezinkampani zabasebenzisi uncane khona.
Ezweni elikhuluma iSpanishi, lapho izinkampani eziningi ezintsha zisebenza ngesabelomali esincane kakhulu kunezinkampani ezisebenza nazo e-United States, ingcindezi yokuncintisana iyabasiza. Uma amamodeli anamandla futhi evulekile etholakala, kulapho amaqembu obuchwepheshe kuzodingeka akhethe khona ngokusekelwe entengo, ukuthobela imithetho, kanye nokusetshenziswa kwawo.hhayi nje kuphela okuvela kumkhiqizo ongemuva kwe-API.
Ngesikhathi esifanayo, i-DeepSeek iyazi ukuthi ukubheja kwayo akunazo izinselelo: iningi lezilinganiso kanye nokuqhathanisa kuvela kumadokhumenti ayo noma ezivivinyweni ezisezingeni lokubuka kuqala, futhi imakethe isalinde ukubona ukuthi amamodeli e-V4 asebenza kanjani uma esetshenziswa kakhulu ezindaweni zokukhiqiza ezidinga kakhulu, kufaka phakathi ezaseYurophu.
Sekukonke, ukufika kwe-DeepSeek V4 kuqinisa umkhuba obekulokhu uthuthuka isikhathi esithile: Amamodeli e-AI asezingeni eliphezulu awaseyona indawo eyingqayizivele yezinkampani ezimbalwa ezinezinhlelo ezivaliwe kanye nesabelomali esikhulu.Ngokuhlanganiswa kwamapharamitha angaphezu kwe-1T, umongo wamathokheni angu-1M, izisindo ezivulekile, kanye nenkulumo egxile ekusebenzeni kahle, inkampani yaseShayina yethula enye indlela izinkampani nabathuthukisi eSpain naseYurophu abangeke bakwazi ukuyishaya indiva ezinhlelweni zabo ezizayo zokwamukela nokuvuselela ingqalasizinda ye-AI.
