- I-Mistral 3 ihlanganisa ndawonye amamodeli ayishumi avuliwe, kusukela kumngcele we-multimodal kuya ochungechungeni oluhlangene lwe-Ministral 3.
- I-Mixture of Expert architecture inika amandla ukunemba okuphezulu ngokusetshenziswa kwamandla okuphansi kanye nokuthunyelwa konqenqemeni okusebenzayo.
- Amamodeli amancane angasebenza angaxhunyiwe ku-inthanethi ku-GPU eyodwa noma amadivayisi anezinsiza eziphansi, okuqinisa ubukhosi bedijithali.
- I-Europe iyazuza ku-AI ngenxa yendlela evulelekile ye-Mistral kanye nokusebenzisana kwayo nezinhlangano zomphakathi nezinkampani.
Ukuqala kweFrance I-Mistral AI Izibeke esikhungweni senkulumompikiswano mayelana nobuhlakani bokwenziwa eYurophu ne Ukwethulwa kwe-Mistral 3Umndeni omusha wamamodeli avuliwe adizayinelwe ukusebenza kuzo zombili izikhungo zedatha ezinkulu namadivayisi anezinsiza ezilinganiselwe kakhulu. Ikude nokungenela umjaho oyimpumputhe wosayizi wamamodeli, inkampani Ikhuthaza ubuhlakani obusabalalisiwe obungasetshenziswa lapho kudingeka khona.: emafini, emaphethelweni, noma ngaphandle kokuxhumeka kwe-inthanethi.
Lelisu libeka I-Mistral njengenye yezindlela ezimbalwa zase-Europe ezikwazi ukumelana neziqhwaga ezifana ne-OpenAI, i-Google noma i-Anthropic, futhi unikele ezinye izindlela ze-ChatGPTKodwa ngombono ohlukile: amamodeli anesisindo esivulekile ngaphansi kwelayisensi evumelekileivumelane nezidingo zezinkampani kanye nokuphathwa komphakathi, futhi ngokugxila okuqinile ezilimini zase-Europe kanye nokuthunyelwa okuzimele ngaphakathi kwezwekazi.
Iyini i-Mistral 3 futhi kungani ibalulekile?
Umndeni I-Mistral 3 Yakhiwa ngu amamodeli ayishumi esisindo avulekile ikhishwe ngaphansi kwe-Apache License 2.0Lokhu kuvumela ukusetshenziswa kwayo kwezohwebo ngaphandle kwemikhawulo. Kufaka phakathi imodeli yohlobo lwe-Frontier. I-Mistral Enkulu 3kanye nomugqa wamamodeli ahlangene ngaphansi komkhiqizo Ungqongqoshe 3eziza ngosayizi abathathu abalinganiselwe (amapharamitha angu-14.000, 8.000 kanye nezigidi ezingu-3.000) kanye nezinhlobonhlobo eziningana kuye ngohlobo lomsebenzi.
Ukuqamba okusha okubalulekile ukuthi imodeli enkulu ayikhawulelwe embhalweni: I-Mistral Large 3 inezinhlobo eziningi futhi inezilimi eziningiIyakwazi ukusebenza ngombhalo nezithombe ngaphakathi kwesakhiwo esifanayo futhi inikeza ukusekelwa okuqinile kwezilimi zaseYurophu. Ngokungafani nezinye izindlela ezihlanganisa amamodeli olimi nombono ngokwehlukana, lena incike ohlelweni olulodwa oluhlanganisiwe olungahlaziya imibhalo emikhulu, luqonde izithombe, futhi lusebenze njengomsizi othuthukile wemisebenzi eyinkimbinkimbi.
Ngesikhathi esifanayo, uchungechunge Ungqongqoshe 3 Idizayinelwe ukusebenza ezimeni lapho ukufinyelela kwamafu kunqunyelwe noma kungekho khona. Lawa mamodeli angasebenza kumadivayisi anento encane I-4 GB yememori noma ku-GPU eyodwa, evula umnyango wokusetshenziswa kwayo amakhompyutha aphathekayo, omakhalekhukhwini, amarobhothi, ama-drones, noma amasistimu ashumekiwe ngaphandle kokuncika ekuxhumekeni kwe-inthanethi okuqhubekayo noma abahlinzeki bangaphandle.
Okwe-ecosystem yaseYurophu, lapho ingxoxo mayelana ubukhosi bedijithali nokulawulwa kwedatha Le nhlanganisela yemodeli yomngcele ovulekile kanye namamodeli angasindi angasebenziseka endaweni ikhona kakhulu futhi ifaneleka ngokukhethekile, kokubili ezinkampanini ezizimele kanye nabaphathi bomphakathi abafuna ezinye izindlela zokusebenzisa izinkundla ezinkulu zase-US naseShayina.
I-Architecture, Ingxube Yochwepheshe, kanye Nendlela Yobuchwepheshe
Inhliziyo yobuchwepheshe ye I-Mistral Enkulu 3 kuyisakhiwo se Ingxube Yochwepheshe (MoE), umklamo lapho imodeli Ine "ochwepheshe" abaningi bangaphakathi.kodwa yenza ingxenye yawo kuphela isebenze ukuze icubungule ithokheni ngayinyeEmpeleni, isistimu ibamba 41.000 billion amapharamitha asebenzayo ngaphandle kwengqikithi ye 675.000 billionLokhu kuvumela ukuhlanganisa amandla okucabanga aphezulu namandla alawulwa kakhulu kanye nokusetshenziswa kwekhompyutha kunemodeli eminyene efanayo.
Lesi sakhiwo, sihlanganiswe ne-a ifasitela lomongo lamathokheni afinyelela ku-256.000Lokhu kuvumela i-Mistral Large 3 ukuthi icubungule imiqulu emikhulu yolwazi, njengezinkontileka ezinde, imibhalo yezobuchwepheshe, noma izisekelo zolwazi ezinkulu zenkampani. Imodeli ihloselwe ezimweni zokusetshenziswa ezifana ukuhlaziywa kwamadokhumenti, usizo lokuhlela, ukudalwa kokuqukethwe, ama-ejenti we-AI, nokuzenzakalelayo kokuhamba komsebenzi.
Ngokuhambisanayo, amamodeli Ungqongqoshe 3 Ahlinzekwa ngezinhlobonhlobo ezintathu eziyinhloko: Isisekelo (imodeli ejwayelekile eqeqeshwe kusengaphambili), Yala (elungiselelwe ingxoxo nemisebenzi yomsizi) kanye Ukubonisana (Kulungiselwe ukucabanga okunengqondo nokuhlaziya okujulile). Zonke izinguqulo zisekela umbono futhi baphatha izimo ezibanzi—phakathi kwamathokheni angu-128K no-256K—, kuyilapho begcina ukuhambisana nezilimi eziningi.
Umbono oyisisekelo, njengoba uchazwe umsunguli kanye nososayensi omkhulu u-Guillaume Lample, ukuthi "ngaphezu kuka-90%" wamacala okusebenzisa amabhizinisi, Imodeli encane, ecushwe kahle yanele. futhi, ngaphezu kwalokho, isebenza kahle kakhulu. Ngokusebenzisa amasu afana nokusetshenziswa kwe idatha yokwenziwa yemisebenzi ethileInkampani ithi lawa mamodeli angasondela noma adlule izinketho ezinkulu, ezivaliwe ezinhlelweni eziqondile, kuyilapho enciphisa izindleko, ukubambezeleka, kanye nezingozi zobumfihlo.
Yonke le ecosystem ihlanganiswe nohlu olubanzi lwemikhiqizo yenkampani: kusuka I-Mistral Agents APInezixhumi zokwenza ikhodi, usesho lwewebhu, noma ukukhiqiza isithombe, kuze kufike Ikhodi ye-Mistral Ukuze uthole usizo lomhleli, imodeli yokucabanga Ngobuciko kanye nesikhulumi Isitudiyo se-AI ukukhipha izinhlelo zokusebenza, ukuphatha izibalo, nokugcina amalogi okusetshenziswa.
Ukusebenzisana ne-NVIDIA kanye nokuthunyelwa ku-supercomputing kanye ne-edge computing
Okugqamile kokwethulwa wumbimbi phakathi I-Mistral AI ne-NVIDIA, ebeka i-Mistral 3 njengomndeni wamamodeli ashunwe kahle kumasistimu e-supercomputing kanye nezinkundla ezisemaphethelweni zomkhiqizi waseMelika. I-Mistral Enkulu 3kuhlanganiswe nengqalasizinda efana I-NVIDIA GB200 NVL72, ngokusho kwe-NVIDIA ukuthuthukiswa kokusebenza okufika izikhathi eziyishumi uma kuqhathaniswa nesizukulwane sangaphambilini esisekelwe kuma-H200 GPUs, sisebenzisa ithuba lokuhambisana okuthuthukile, inkumbulo eyabiwe nge-NVLink, namafomethi ezinombolo athuthukisiwe afana I-NVFP4.
Umsebenzi wokuhlanganyela awumi ku-hardware yezinga eliphezulu. Uchungechunge Ungqongqoshe 3 Ilungiselelwe ukusebenza ngokushesha ezindaweni ezifana Ama-PC namalaptop anama-RTX GPU, amadivaysi e-Jetson, namapulatifomu asemaphethelweniukwenza kube lula ukucatshangwa kwasendaweni kuzimboni, amarobhothi, noma izimo zabathengi. Popular izinhlaka ezifana I-Llama.cpp kanye no-Ollama Aguqulelwe ukuze asebenzise lawa mamodeli, okwenza kube lula ukuthunyelwa kwawo ngabathuthukisi namaqembu e-IT.
Ngaphezu kwalokho, ukuhlanganiswa ne-ecosystem I-NVIDIA NeMo —okuhlanganisa amathuluzi afana noMklami Wedatha, I-Guardrails, ne-Agent Toolkit— kwenza izinkampani zikwazi ukusebenza ukulungisa kahle, ukulawula ukuphepha, i-orchestration ye-ejenti, nomklamo wedatha esekelwe Mistral 3. Ngesikhathi esifanayo, izinjini inference ezifana I-TensorRT-LLM, i-SGlang ne-vLLM ukunciphisa izindleko ngethokheni ngayinye nokuthuthukisa ukusebenza kahle kwamandla.
Amamodeli we-Mistral 3 manje ayatholakala ezitolo ezinkulu abahlinzeki befu nezinqolobane ezivulekilefuthi bazofika futhi ngesimo se I-NIM microservices ngaphakathi kwekhathalogi ye-NVIDIA, okuthile okuthakazelisa ngokukhethekile ezinkampanini zaseYurophu esezivele zisebenza ezitakini zalo mkhiqizi futhi ezifuna ukusebenzisa i-AI ekhiqizayo enokulawula okukhulu ekusetshenzisweni.
Lonke lolu hlaka luvumela i-Mistral 3 ukuthi iphile kokubili ezikhungweni zedatha ezinkulu nakumadivayisi asemaphethelweni, igcizelela ukulandisa kwayo kwe itholakala yonke indawo futhi i-AI esabalalisiwe, incike kancane kumasevisi akude futhi ijwayelane nezidingo ezithile zeklayenti ngalinye.
Amamodeli amancane, ukusetshenziswa okungaxhunyiwe ku-inthanethi, namakesi okusetshenziswa asemaphethelweni
Enye yezinsika zenkulumo kaMistral yileyo Izinhlelo zokusebenza eziningi zomhlaba wangempela azidingi imodeli enkulu kakhulu engenzeka.kodwa ehambisana kahle necala lokusebenzisa futhi ingacushwa kahle nedatha ethile. Yilapho kungena khona amamodeli ayisishiyagalolunye ochungechungeni. Ungqongqoshe 3kuminyene, ukusebenza okuphezulu, futhi kutholakala ngosayizi abahlukene nokuhlukahluka ukuze kuhambisane nezindleko, isivinini, noma izidingo zomthamo.
Lawa mamodeli aklanyelwe ukusebenza kuwo i-GPU eyodwa noma ihadiwe enesizothaLokhu kuvumela ukusetshenziswa kwendawo kumaseva angaphakathi endlini, amakhompyutha aphathekayo, amarobhothi ezimboni, noma amadivayisi asebenza ezindaweni ezikude. Ezinkampanini eziphethe ulwazi olubucayi—kusuka kubakhiqizi kuya ezikhungweni zezezimali noma izikhungo zikahulumeni—ikhono lokuqhuba i-AI ngaphakathi kwengqalasizinda yazo, ngaphandle kokuthumela idatha emafini, kuyinzuzo enkulu.
Inkampani icaphuna izibonelo ezifana nalezi Amarobhothi efekthri ahlaziya idatha yezinzwa ngesikhathi sangempela ngaphandle kokuxhumeka kwe-inthanethi, ama-drones ezimo eziphuthumayo nezokutakula, izimoto ezisebenza ngokugcwele abasizi be-AI ezindaweni ezingenayo ukufakwa. noma amathuluzi okufundisa anikeza usizo olungaxhunyiwe ku-inthanethi kubafundi. Ngokucubungula idatha ngqo kudivayisi, i ubumfihlo kanye nokulawulwa kolwazi yabasebenzisi.
U-Lample ugcizelela ukuthi ukufinyeleleka kuyingxenye eyinhloko yomsebenzi we-Mistral: kukhona Izigidigidi zabantu abanamaselula noma amalaptop kodwa abangenakho ukufinyelela kwe-inthanethi okuthembekileezingazuza kumamodeli akwazi ukugijima endaweni. Ngale ndlela, inkampani izama ukuqeda umbono wokuthi i-AI ethuthukisiwe kufanele ihlale iboshelwe ezikhungweni ezinkulu zedatha ezilawulwa yiqembu elincane lezinkampani.
Ngokuhambisanayo, iMistral isiqalile ukusebenzisana nabalingani bamazwe ngamazwe endaweni eyaziwa ngokuthi I-AI yomzimbaPhakathi kokubambisana okukhulunywe ngakho yi-ejensi ye-HTX yesayensi nobuchwepheshe yaseSingapore yamarobhothi, i-cybersecurity, nezinhlelo zokuvikela umlilo; kanye nesiJalimane I-Helsing, egxile ekuvikelweni, ngamamodeli esenzo solimi lombono wama-drones; kanye nabakhiqizi bezimoto abafuna Abasizi be-AI ekamelweni esebenza kahle futhi elawulekayo.
Umthelela eYurophu: ubukhosi bedijithali kanye nesistimu yezemvelo yomphakathi neyangasese
Ngale kwezici zobuchwepheshe, iMistral isiphenduke ibhentshimakhi kunkulumompikiswano mayelana Ubukhosi bedijithali eYurophuNakuba inkampani izichaza ngokuthi "ukubambisana kwe-transatlantic" -namaqembu kanye nokuqeqeshwa okuyimodeli okusatshalaliswa phakathi kweYurophu ne-United States-, ukuzibophezela kwayo ekuvuleni amamodeli ngokusekelwa okuqinile kwezilimi zaseYurophu kuye kwamukelwe kahle yizikhungo zomphakathi ezwenikazi.
Inkampani ivale izivumelwano nayo ibutho laseFrance, isikhungo saseFrance esiqasha umphakathi, uhulumeni waseLuxembourg, nezinye izinhlangano zaseYurophu unentshisekelo yokuthumela i-AI ngaphansi kwezinhlaka eziqinile zokulawula kanye nokugcina ukulawulwa kwedatha ngaphakathi kwe-EU. Ngokuhambisanayo, i-European Commission yethule a isu lokuthuthukisa amathuluzi e-AI yaseYurophu eziqinisa ukuncintisana kwezimboni ngaphandle kokudela ukuphepha nokuqina.
Umongo we-geopolitical nawo ucindezela isifunda ukuthi sisabele. Kuyaqashelwa lokho IYurophu isilele ngemuva kwe-United States neChina Emjahweni wamamodeli esizukulwane esilandelayo, ngenkathi emazweni afana ne-China izindlela ezivulekile ezivulekile ezifana ne-DeepSeek, i-Alibaba, ne-Kimi ziyavela futhi ziqala ukuncintisana nezixazululo ezifana ne-ChatGPT emisebenzini ethile, i-Mistral izama ukugcwalisa ingxenye yalelo gebe ngamamodeli avulekile, aguquguqukayo ahambisana nezidingo zokulawula zaseYurophu.
Ngokwezimali, ukuqaliswa kukhuphuke cishe AmaRandi ayizigidigidi eziyi-2.700 futhi ihambe ngaphakathi kokulinganisa eduze 14.000 billionLezi zibalo ziphansi kakhulu kunalezo zemidondoshiya njenge-OpenAI noma i-Anthropic, kodwa zibalulekile ku-ecosystem yaseYurophu. Ingxenye enkulu yemodeli yebhizinisi ibandakanya ukunikela, ngale kwezisindo ezivulekile, amasevisi okwenza ngokwezifiso, amathuluzi okuthunyelwa, nemikhiqizo yebhizinisi njenge-Mistral Agents API noma i-Le Chat suite enokuhlanganiswa kwezinkampani.
Ukuma kucacile: ukuba a umhlinzeki wengqalasizinda ye-AI evulekile futhi eguquguqukayo lokho kuvumela izinkampani zaseYurophu (nezinye izifunda) ukuthi zisungule izinto ezintsha ngaphandle kokuncika ngokuphelele ezinkundleni zokuxhumana zase-US, kuyilapho zigcina ukulawula ukuthi amamodeli asetshenziswa kuphi futhi kanjani, kanye nokwenza lula ukuhlanganiswa ngamathuluzi asevele asetshenziswa ezinhlelweni zawo.
Inkulumo-mpikiswano ngokuvuleleka kwangempela nezinselele ezisalindile
Naphezu kwesasasa elikhiqizwa yi-Mistral 3 engxenyeni yomphakathi wezobuchwepheshe, akukho ukushoda kwamazwi abucayi lowo mbuzo. ukuthi la mamodeli angacatshangelwa kangakanani ngempela "umthombo ovulekile"Inkampani ikhethe indlela yokwenza isisindo esivulekileIkhulula izisindo ezizosetshenziswa futhi zijwayelwe, kodwa hhayi yonke imininingwane mayelana nedatha yokuqeqeshwa nezinqubo zangaphakathi ezidingekayo ukukhiqiza kabusha imodeli kusukela ekuqaleni.
Abacwaningi abafana no Andreas Liesenfeld, umsunguli we-European Open Source AI Index, Baveza ukuthi ibhodlela elikhulu le-AI eYurophu akukhona nje ukufinyelela kumamodeli, kodwa ku idatha yokuqeqeshwa kwezinga elikhuluKusukela kulowo mbono, i-Mistral 3 inomthelela thuthukisa uhla lwamamodeli asebenzisekayoKodwa-ke, ayixazululi ngokugcwele inkinga ewumsuka ye-ecosystem yaseYurophu eqhubeka nokuzabalaza ukukhiqiza nokwabelana ngamadathasethi amakhulu wekhwalithi ephezulu.
I-Mistral ngokwayo iyavuma ukuthi amamodeli ayo avulekile "angemuva kancane" kwezixazululo ezivaliwe ezithuthuke kakhulu, kodwa Ugcizelela ukuthi igebe liyancipha ngokushesha. nokuthi iphuzu eliyinhloko liwukuthi isilinganiso sezindleko nenzuzoUma imodeli enamandla kancane ingafakwa ngezindleko eziphansi, ilungiselwe umsebenzi othile, futhi isebenze eduze nomsebenzisi, Lokhu kungase kuthakazelise ezinkampanini eziningi kunemodeli ephezulu engafinyelelwa kuphela nge-API eyirimothi.
Noma kunjalo, izinselelo zisekhona: kusukela ku- ukuncintisana kwamazwe ngamazwe okushubile Lokhu kudlulela esidingweni sokuqinisekisa ukuvikeleka, ukulandeleka, nokuthobela imithetho ezimeni ezifana nokunakekelwa kwezempilo, ezezimali, kanye nohulumeni. Ibhalansi phakathi kokuvuleleka, ukulawula, nokuzibophezela kuzoqhubeka nokuqondisa iMistral nabanye abadlali baseYurophu eminyakeni ezayo.
Ukwethulwa kwe I-Mistral 3 Iqinisa umbono wokuthi i-AI esezingeni eliphezulu akudingeki ukuthi ikhawulelwe kumamodeli amakhulu, avaliwe.futhi inikeza i-Europe - nanoma iyiphi inhlangano eyazisa ubukhosi bezobuchwepheshe - iphalethi yamathuluzi avulekile ahlanganisa imodeli ye-multimodal frontier nohlu lwamamodeli angasindi akwazi ukusebenza onqenqemeni, ungaxhunyiwe ku-inthanethi, futhi nezinga lokwenza ngendlela oyifisayo okunzima ukulifanisa ngezinkundla zobunikazi kuphela.
Ngingumshisekeli wezobuchwepheshe oguqule izintshisekelo zakhe "ze-geek" zaba umsebenzi. Ngichithe iminyaka engaphezu kwengu-10 yempilo yami ngisebenzisa ubuchwepheshe obusezingeni eliphezulu kanye nokukitaza ngazo zonke izinhlobo zezinhlelo ngenxa yelukuluku lokufuna ukwazi. Manje sengiqeqeshelwe ubuchwepheshe be-computer nemidlalo yama-video. Lokhu kungenxa yokuthi sekuphele iminyaka engaphezu kwengu-5 ngisebenza ngokubhalela amawebhusayithi ahlukahlukene ezobuchwepheshe nemidlalo yevidiyo, ngenza izindatshana ezifuna ukukunikeza imininingwane oyidingayo ngolimi oluqondakala yiwo wonke umuntu.
Uma unemibuzo, ulwazi lwami lusukela kuyo yonke into ehlobene nesistimu yokusebenza ye-Windows kanye ne-Android yomakhalekhukhwini. Futhi ukuzibophezela kwami kuwe, ngihlala ngizimisele ukuchitha amaminithi ambalwa futhi ngikusize uxazulule noma yimiphi imibuzo ongase ube nayo kulo mhlaba we-inthanethi.