Kwenzekeni
I-DeepSeek ishicilele i-DeepSeek-V4-Flash-Vision-Exp ku-Hugging Face njengemodeli yayo yokuqala yokuhlola ye-multimodal emndenini we-DeepSeek-V4. Ikhadi eliyimodeli lithi lengeza amamojula abukwayo kanye nokuqeqeshwa okuqhubekayo ekwakhiweni kwe-DeepSeek-V4-Flash, okunezinzuzo ezibikiwe emisebenzini ye-ejenti ye-multimodal kuyilapho kugcinwa ukusebenza okuqhathaniswayo emisebenzini ye-ejenti yombhalo kuphela.
Irekhodi le-Hugging Face lihlonza i-DeepSeek-V4-Flash-Vision-Exp njengemodeli yombhalo-umbhalo-umbhalo usebenzisa i-Transformers. Ikhadi layo eliyimodeli liyichaza njengemodeli yokuqala yokuhlola ye-multimodal ye-DeepSeek emndenini we-V4. Umklamo oshiwo uhlanganisa ukwakheka kwe-DeepSeek-V4-Flash namamojula abonakalayo kanye nokuqeqeshwa okuqhubekayo okuhloselwe ukuvula amandla okuqonda okubukwayo. Indawo yokugcina yakhiwe ngo-Agasti 31, 2026, futhi irekhodi libonisa isibuyekezo esilandelayo ngoSepthemba 1.
Ikhadi eliyimodeli libika ukuthuthuka ngaphezu kwe-DeepSeek-V4-Flash-0731 ekuhloleni okumbalwa kwe-ejenti ye-multimodal. Ibala amaphuzu e-ApexBench Pass@1 angu-36.5 uma kuqhathaniswa nama-26.2 emodeli yangaphambili, amaphuzu e-Agents’ Last Exam angu-27.3 uma kuqhathaniswa nama-25.2, amaphuzu e-Chartography angu-64.3, kanye ne-ZeroBench Pass@5 amaphuzu angu-35.0. Ikhadi liphinde liqhathanise imodeli entsha ne-Opus-4.8, eliyibala ku-39.4 ku-ApexBench, 25.7 ku-Agents’ Last Exam, 65.0 ku-Chartography, kanye no-34.0 ku-ZeroBench.
Ngemisebenzi yomenzeli wombhalo, ikhadi eliyimodeli libika 83.9 ku-Terminal Bench 2.1, 57.7 ku-NL2Repo, 75.3 ku-Cybergym, 59.3 ku-DeepSWE, 75.9 ku-Toolathlon-Verified, 63.6 ku-DSBench-Hard, kanye no-25.7 ku-Auto Public. Ikhadi lithi imodeli entsha igcina ukusebenza kwe-ejenti yombhalo kuphela okuqhathanisekayo ne-DeepSeek-V4-Flash-0731, kuyilapho iphawula ukuthi imodeli yangaphambili yayiziba izici ze-multimodal ku-ApexBench kanye nokokufaka kwe-Agents' Last Exam.
Inqolobane ihlanganisa amafayela ethokheni, izinkomba zombhalo osheshayo, kanye nokusetshenziswa okuncane kwe-PyTorch okuchazayo okuhlanganisa isishumeki sombono nokuqondanisa, ukunaka kwe-DFlash, izingxenye ezixubile zochwepheshe, i-Hyper-Connections, kanye nendlela eya phambili ye-DSpark. Ikhadi eliyimodeli linikeza imiyalelo ye-Transformers, i-vLLM, i-Docker, i-SGlang, ne-Docker Model Runner. Isibonelo sayo se-SGlang sicacisa ukufana kwe-tensor kwezine kanye ne-DSpark . Imethadatha ebonakalayo ibala amapharamitha ayizigidi eziyizinkulungwane ezingama-305, futhi indawo yokugcina ilayisensi ngaphansi kwe-MIT.
Imininingwane yomthombo: huggingface.co ↗
Kungani kubalulekile
Ukukhishwa kunikeza abacwaningi nabathuthukisi ukufinyelela kumodeli enkulu, enelayisensi ye-MIT edizayinelwe ukuxhumana kwesithombe nombhalo kanye nokugeleza komsebenzi kwe-ejenti. Ukusebenza kwayo okungokoqobo kuhlala kungaqinisekile ngenxa yokuthi ukuhlaziya okubikiwe kuvela ekhadini eliyimodeli, elimakwe ngokuthi aliqinisekisiwe, futhi imodeli ayisetshenziswanga umhlinzeki wemibono.
Lokhu kukhishwa kubalulekile ngoba umbono uyinhloso eqondile yemodeli esikhundleni sesici sengozi. Isikhundla sekhadi lemodeli DeepSeek-V4-Flash-Vision-Exp samakhono e-ejenti ye-multimodal, okusho imisebenzi edinga isistimu ye-AI ukuze ihumushe okokufaka okubonakalayo okuhambisana nolimi nokusebenza ngaphakathi kohlaka lokuhlola. Lokho kunweba uhlobo lokokufaka okuhloselwe ukukusingatha umndeni we-V4-Flash, nakuba umthombo ungaqinisekisi ukuthi imodeli isebenza kahle kangakanani emikhiqizweni evamile noma kuzilungiselelo eziphezulu.
Inqolobane yenza imodeli isebenziseke kakhulu konjiniyela abafuna ukuhlola, ukujwayela, noma ukusebenzisa isistimu ye-multimodal yokuhlola. Ilayisense ye-MIT, izinkomba zombhalo osheshayo, amafayela we-tokenizer, ikhodi ye-inference, nezibonelo zinikeza izinto zokuqalisa eziningi kunesimemezelo somkhiqizo kuphela. Ngesikhathi esifanayo, umthombo uthi ama-shards amamodeli amakhulu achazwa yinkomba futhi awaphindaphindeki ngaphakathi kokuphuma komthombo okusetshenziselwa ukuhlanganisa indawo yokugcina. Usayizi wemodeli nokucushwa okufakwe ohlwini kwe-GPU SGLang eningi kubonisa ukuthi ukusetshenziswa kungase kudinge, kodwa umthombo awunikezi izindleko zangempela zehadiwe, ukubambezeleka, noma izidingo zememori.
Imiphumela ebikiwe iphakamisa ukuhwebelana okuthile: I-DeepSeek ifuna izinzuzo ezinkulu kumabhentshimakhi e-ejenti ye-multimodal ngaphandle kokuyeka ukusebenza kwe-ejenti yombhalo okuqhathanisekayo. Lezo zimangalo kufanele zithathwe njengemiphumela yekhadi eliyimodeli esikhundleni samaqiniso amisiwe azimele. Amarekhodi okuhlola akhomba ikhadi eliyimodeli njengomthombo walo futhi amaka imiphumela njengengaqinisekisiwe. Ukuqhathanisa nakho akuphelele ezindaweni: ezinye izikolo zamamodeli angaphambilini amamodeli amaningi zimakwe ngenothi elichaza ukuthi imodeli iyaziba izici ezibonakalayo, kuyilapho amanye amaseli ebhentshimakhi angenayo imiphumela yemodeli yangaphambili.
Ngakho-ke umthelela womphakathi uncike ekuqinisekiseni nasekusebenziseni. Uma ukuhlola okuzimele kuqinisekisa izinzuzo ezibikiwe zomenzeli obonakalayo futhi ukusetshenziswa kungaqhutshwa abacwaningi abaningi, ukukhishwa kungase kunwebe ukufinyelela kumodeli ekwaziyo evulekile yokuhlolwa kwesithombe nombhalo. Umthombo awusunguli ukuqeqeshwa kwedatha, ukuhlolwa kokuphepha, ukuziphatha okwenqaba, ukuvikelwa kobumfihlo, usekelo lwezentengiso, noma ukufaneleka kwezinqumo ezilandelanayo. Lokho kweqiwe kukhawulela lokho okungaphethwa ngokuzibophezela kusukela ekukhululweni kukodwa.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Imibuzo eyinhloko ukuthi ingabe ukuhlola okuzimele kukhiqiza kabusha imiphumela ye-DeepSeek, ukuthi imodeli idinga izingxenyekazi ezingakanani zehadiwe nobunjiniyela ekusebenzeni, nokuthi ukuqaliswa kokuhlola kwendawo yokugcina kuba lula yini ukukusebenzisa. Abasebenzisi kufanele futhi babheke ulwazi olugcwele mayelana nedatha, ukuhlolwa kokuphepha, nokusebenza ngaphandle kwamabhentshimakhi asohlwini.
Ukutholakala umbuzo osebenzayo osheshayo. Ikhasi le-Hugging Face lithi imodeli ayisetshenziswanga yinoma yimuphi umhlinzeki we-inference, futhi isifinyezo somthombo sibonisa okulandwayo okuyiziro. Abasebenzisi esikhundleni salokho baqondiswe emizileni yasendaweni noma eziphethwe yona njengeTransformers, vLLM, SGlang, Docker, kanye ne-Docker Model Runner. Izibuyekezo zesikhathi esizayo zingacacisa ukuthi ukufinyelela okusingathiwe kuyatholakala yini, noma ngabe indlela encane yokukhomba iphelile, nokuthi yiziphi izilungiselelo zezingxenyekazi zekhompuyutha ezingokoqobo ngale kwesibonelo se-tensor-parallel-four.
Ukuphindaphinda okuzimele kufanele kugxile ezimangalweni ze-multimodal nasekuqhathaniseni okulungile nemodeli yangaphambili. Abahloli bazodinga ukulandisa ngenothi lomthombo ukuthi i-DeepSeek-V4-Flash-0731 ishaye indiva izici ezibonakalayo ekuhlolweni okubili okusohlwini. Kufanele futhi bahlole isimo esingaqinisekisiwe semiphumela yemodeli yekhadi, babike ukwaziswa okunembile nezilungiselelo zokuhlanganisa, futhi bahlole ukuthi ingabe ukusebenza kubamba kuzo zonke izithombe, izilimi, imisebenzi, kanye nezigameko zokwehluleka ezingamelwe ithebula elishicilelwe.
Isimo sokuhlola sokukhishwa sigunyaza ukunakwa kwezinguquko zobunjiniyela. Inqolobane ihlukanisa ukufometha osheshayo kusuka ku-PyTorch inference, isekela kokubili amabhulokhi wokuqukethwe wesitayela se-OpenAI se-JSON kanye nombhalo ohlangene, kanye nokuguqulwa kwendawo yokuhlola amadokhumenti. Izibuyekezo kulezo zingxenye zingathinta ukukhiqizwa kabusha nokusetshenziswa. Umthombo awusho ukuthi izixhumanisi ziqine kangakanani, ukuthi izisindo noma ikhodi izoshintsha kangaki, noma ukuthi zonke izindlela zokuphakela ezibhalwe phansi zisekela izici zombono ngokulinganayo.
Ulwazi lwezokuphepha nokubusa lungenye enkulu engaziwa. Umthombo uchaza izakhiwo, ukusetshenziswa, amabhentshimakhi, nokunikezwa kwelayisensi kodwa awunikezi ukuhlolwa kokuphepha noma imikhawulo yokuqonda isithombe. Ngaphambi kokusebenzisa imodeli enezithombe ezibucayi noma ukugeleza komsebenzi okulandelanayo, izinhlangano zizodinga ubufakazi mayelana nokuphathwa kwedatha, amaphutha okubukwayo, ukuphepha, ukumelana nokusetshenziswa kabi, kanye nokugadwa komuntu. Azikho kulezo zakhiwo ezingathathwa kusukela kuzikolo zokulinganiswa noma ilayisensi ye-MIT.