- I-Mistral 3 izisa iimodeli ezilishumi ezivulekileyo, ukusuka kumda we-multimodal ukuya kwi-compact Ministral 3 series.
- I-Mixture of Expert architecture yenza ukuchaneka okuphezulu kunye nokusetyenziswa kwamandla aphantsi kunye nokuhanjiswa okusebenzayo komda.
- Iimodeli ezincinci zinokusebenza ngaphandle kweintanethi kwi-GPU enye okanye izixhobo ezinobuncwane obuphantsi, ukuqinisa ulongamo lwedijithali.
- IYurophu izuza umhlaba kwi-AI ngenxa yendlela evulelekileyo yeMistral kunye nentsebenziswano namaqumrhu karhulumente kunye neenkampani.
Ukuqala kweFransi I-Mistral AI Iye yazibeka embindini wengxoxo ngobukrelekrele bokwenziwa eYurophu kunye ne Ukuqaliswa kweMistral 3Intsapho entsha yeemodeli ezivulekileyo ezenzelwe ukusebenza kuzo zombini amaziko amakhulu edatha kunye nezixhobo ezinobuncwane obuncinci kakhulu. Kude nokungenela ugqatso oluyimfama lobungakanani bemodeli, inkampani Ixhasa ukusasazwa kobuntlola obunokuphunyezwa naphi na apho kuyimfuneko.: efini, emphethweni, okanye ngaphandle koqhagamshelo lwe-intanethi.
Esi sicwangciso sibeka IMistral njengenye yeendlela ezimbalwa zaseYurophu ezikwaziyo ukumelana nezigebenga ezifana ne-OpenAI, uGoogle okanye iAnthropic., kwaye unikeze ezinye iindlela ze-ChatGPTKodwa ngokwembono eyahlukileyo: iimodeli ezinobunzima obuvulekileyo phantsi kwelayisensi evumelekileyoukulungelelaniswa neemfuno zeenkampani kunye nolawulo lukarhulumente, kwaye ngokugxila ngamandla kwiilwimi zaseYurophu kunye nokuthunyelwa okuzimeleyo kwilizwekazi.
Yintoni iMistral 3 kwaye kutheni ibalulekile?
Usapho IMistral 3 Yenziwa ngu iimodeli zobunzima ezilishumi ezivulekileyo ikhutshwe phantsi kwe-Apache License 2.0Oku kuvumela usetyenziso lwayo lwentengiso ngaphandle kwezithintelo. Ibandakanya imodeli yohlobo lweFrontier. IMistral Enkulu 3kunye nomgca weemodeli ezidibeneyo phantsi kwe-brand UMphathiswa 3eziza ngokweesayizi ezintathu eziqikelelwayo (14.000, 8.000 kunye ne-3.000 yezigidi zeeparamitha) kunye neentlobo ezininzi ezixhomekeke kuhlobo lomsebenzi.
Utshintsho oluphambili kukuba imodeli enkulu ayiphelelanga kwisicatshulwa: I-Mistral Large 3 ineendlela ezininzi kunye neelwimi ezininziIyakwazi ukusebenza ngokubhaliweyo kunye nemifanekiso ngaphakathi kwesakhiwo esifanayo kwaye inikezela ngenkxaso enamandla kwiilwimi zaseYurophu. Ngokungafaniyo nezinye iindlela ezidibanisa ulwimi kunye neemodeli zombono ngokwahlukileyo, lo uthembele kwinkqubo enye edibeneyo ekwazi ukuhlalutya amaxwebhu amakhulu, ukuqonda imifanekiso, kwaye enze njengomncedisi ophambili wemisebenzi enzima.
Ngexesha elifanayo, uchungechunge UMphathiswa 3 Yenzelwe ukusebenza kwiimeko apho ukufikelela kwifu kuthintelwe okanye kungabikho. Ezi modeli zinokusebenza kwizixhobo ezinobuncinci I-4 GB yememori okanye kwi-GPU enye, evula umnyango wokusetyenziswa kwayo iilaptops, iiselfowuni, iirobhothi, iidrone, okanye iinkqubo ezizinzisiweyo ngaphandle kokuxhomekeka kuqhagamshelo lwe-intanethi rhoqo okanye ababoneleli bangaphandle.
Kwi-ecosystem yaseYurophu, apho incoko malunga ulongamo lwedijithali kunye nolawulo lwedatha Le ndibaniselwano yemodeli yemida evulekileyo kunye neemodeli ezinobunzima obulula ezinokubekelwa kwindawo ikhona kakhulu kwaye ifanelekile, zombini iinkampani zabucala kunye nolawulo lukarhulumente olufuna ezinye iindlela kumaqonga amakhulu ase-US kunye namaTshayina.
I-Architecture, uMxube weeNgcali, kunye neNdlela yoBugcisa
Intliziyo yobugcisa ye IMistral Enkulu 3 Uyilo loyilo lwe Umxube weeNgcali (MoE), uyilo apho imodeli Ineengcaphephe ezininzi zangaphakathi.kodwa ivula kuphela inxalenye yazo ukuqhubekekisa umqondiso ngamnyeNgokwesiqhelo, inkqubo iyaphatha I-41.000 yeebhiliyoni zeeparamitha ezisebenzayo malunga iyonke 675.000 izigidiOku kuvumela ukudibanisa amandla okuqiqa aphezulu kunye namandla alawulwayo kunye nokusetyenziswa kwekhompyutha kunemodeli eshinyeneyo elinganayo.
Olu lwakhiwo, ludityaniswe ne ifestile yemeko ukuya kuthi ga kwi-256.000 iithokheniOku kuvumela iMistral Large 3 ukuba isebenze imiqulu emikhulu yolwazi, efana nezivumelwano ezinde, amaxwebhu obugcisa, okanye iziseko zolwazi ezinkulu. Imodeli ilungiselelwe iimeko zokusetyenziswa ezifana uhlalutyo loxwebhu, uncedo lwenkqubo, ukudalwa komxholo, ii-arhente ze-AI, kunye nokuzenzekelayo kokuhamba komsebenzi.
Ngokuhambelanayo, iimodeli UMphathiswa 3 Zibonelelwa ngeentlobo ezintathu eziphambili: isiseko (imodeli eqeqeshwe kwangaphambili), Yalela (ilungiselelwe incoko kunye nemisebenzi yomncedisi) kunye Ukuqiqa (Ulungelelwaniso lwengqiqo kunye nohlalutyo olunzulu). Zonke iinguqulelo zixhasa umbono kwaye baphatha iimeko ezibanzi-phakathi kwe-128K kunye ne-256K amathokheni-, ngelixa begcina ukuhambelana neelwimi ezininzi.
Umbono osisiseko, njengoko uchazwa ngumseki kunye nososayensi oyintloko uGuillaume Lample, kukuba "ngaphezu kwe-90%" yamatyala okusetyenziswa kwamashishini, Imodeli encinci, elungiswe kakuhle yanele. kwaye, ngaphezu koko, ukusebenza kakuhle. Ngokusebenzisa ubuchule obufana nokusetyenziswa kwe idatha yokwenziwa kwemisebenzi ethileInkampani ixoxa ukuba ezi modeli zinokusondela okanye zigqithe ezinkulu, iinketho ezivaliweyo kwizicelo ezithile, ngelixa zinciphisa iindleko, ukubambezeleka, kunye nemingcipheko yabucala.
Yonke le ecosystem idityaniswe noluhlu olubanzi lweemveliso zenkampani: ukusuka Mistral Agents APInezihlanganisi zokwenza ikhowudi, uphendlo lwewebhu, okanye ukuveliswa komfanekiso, ukuya kuthi ga Ikhowudi yeMistral Ngoncedo lomdwelisi wenkqubo, imodeli yokuqiqa Ngobuchule neqonga Isitudiyo se-AI ukuhambisa izicelo, ukulawula uhlalutyo, kunye nokugcina iilogi zokusetyenziswa.
Ukusebenzisana ne-NVIDIA kunye nokuthunyelwa kwi-supercomputing kunye ne-edge computing
Okugqwesileyo kokuqaliswa yimbumbano phakathi IMistral AI kunye neNVIDIA, ebeka i-Mistral 3 njengentsapho yeemodeli ezilungiswe kakuhle kwiinkqubo ze-supercomputing kunye neeplatifomu zomda womenzi waseMelika. IMistral Enkulu 3kudityaniswe neziseko zophuhliso ezifana NVIDIA GB200 NVL72, ngokutsho kweNVIDIA ukuphuculwa komsebenzi ukuya kumaxesha alishumi xa kuthelekiswa nesizukulwana sangaphambili esekwe kwi-H200 GPUs, ukuthatha ithuba lokuhambelana okuphezulu, inkumbulo ekwabelwana ngayo nge-NVLink, kunye neefomathi zamanani eziphuculweyo ezifana NVFP4.
Umsebenzi wentsebenziswano awumi kwi-hardware ephezulu. Uthotho UMphathiswa 3 Ilungiselelwe ukuba iqhube ngokukhawuleza kwiindawo ezifana IiPC kunye neelaptops ezine-RTX GPUs, izixhobo zeJetson, kunye namaqonga omdaukuququzelela ukucingelwa kwendawo kwimizi-mveliso, iirobhothi, okanye iimeko zabathengi. Izikhokelo ezidumileyo ezifana Llama.cpp kunye no-Ollama Ziye zahlengahlengiswa ukuze zisebenzise le mizekelo, eyenza lula ukuthunyelwa kwabo ngabaphuhlisi kunye namaqela e-IT.
Ngaphaya koko, indibaniselwano ne-ecosystem I-NVIDIA NeMo -kubandakanya izixhobo ezifana noMyili weDatha, i-Guardrails, kunye ne-Agent Toolkit- yenza ukuba iinkampani ziqhube ukulungiswa kakuhle, ulawulo lokhuseleko, i-agent orchestration, kunye noyilo lwedatha esekelwe Mistral 3. Kwangaxeshanye, iinjini inference ezifana I-TensorRT-LLM, i-SGlang kunye ne-vLLM ukunciphisa iindleko ngethokheni kunye nokuphucula ukusebenza kakuhle kwamandla.
Iimodeli ezi-3 zeMistral ngoku ziyafumaneka kubathengisi abakhulu ababoneleli belifu kunye noovimba abavulekileyokwaye nabo baya kufika bekwimo ye Iinkonzo ezincinci ze-NIM ngaphakathi kwekhathalogu ye-NVIDIA, into enomdla ngakumbi kwiinkampani zaseYurophu esele zisebenza kwimfumba yalo menzi kwaye zifuna ukwamkela i-AI yokuvelisa ngolawulo olukhulu lokusasazwa.
Sonke esi sikhokelo sivumela iMistral 3 ukuba ihlale kumaziko amakhulu edatha kunye nakwizixhobo ezisekupheleni, igxininisa ibali layo. ngokwenene yonke indawo kwaye isasazwe i-AI, ngaphantsi kokuxhomekeka kwiinkonzo ezikude kwaye zilungelelaniswe ngakumbi kwiimfuno ezithile zomthengi ngamnye.
Iimodeli ezincinci, ukusasazwa ngaphandle kweintanethi, kunye neemeko zokusetyenziswa komda
Enye yeentsika zentetho kaMistral yileyo Uninzi lwezicelo zehlabathi lokwenyani azifuni eyona modeli inkulu inokwenzeka.kodwa enye ehambelana kakuhle nemeko yokusetyenziswa kwaye inokulungiswa kakuhle ngedatha ethile. Kulapho iimodeli ezisithoba kuthotho zingena khona. UMphathiswa 3xinene, ukusebenza okuphezulu, kwaye kufumaneka ngokobukhulu obahlukeneyo kunye nokwahluka ukuhambelana neendleko, isantya, okanye iimfuno zomthamo.
Ezi modeli ziyilelwe ukusebenza kuzo iGPU enye okanye nakwihardware eqhelekileyoOku kuvumela ukuthunyelwa kwasekhaya kwiiseva zangaphakathi, iilaptops, iirobhothi zamashishini, okanye izixhobo ezisebenza kwindawo ezikude. Kwiinkampani eziphethe ulwazi olubucayi-ukusuka kubavelisi ukuya kumaziko emali okanye ii-arhente zikarhulumente-amandla okuqhuba i-AI ngaphakathi kweziseko zabo, ngaphandle kokuthumela idatha kwilifu, inzuzo ebalulekileyo.
Inkampani ikhankanya imizekelo efana nale Iirobhothi zasefektri ezihlalutya idatha ye-sensor ngexesha langempela ngaphandle koqhagamshelo lwe-intanethi, i-drones kwiimeko ezingxamisekileyo kunye nokuhlangula, izithuthi ezinabancedisi be-AI abasebenzayo ngokupheleleyo kwiindawo ngaphandle kokugubungela. okanye izixhobo zemfundo ezibonelela ngoncedo lwangaphandle kweintanethi kubafundi. Ngokucubungula idatha ngqo kwisixhobo, i ubumfihlo kunye nolawulo lolwazi Abasebenzisi.
I-Lample igxininisa ukuba ukufikeleleka yinxalenye engundoqo yomsebenzi weMistral: zikho Iibhiliyoni zabantu abaneefowuni eziphathwayo okanye iilaptops kodwa bengenakho ukufikelela okuthembekileyo kwi-intanethienokuzuza kwiimodeli ezikwaziyo ukubaleka ekuhlaleni. Ngale ndlela, inkampani izama ukugxotha ingcamango yokuba i-AI ephezulu kufuneka ihlale ibophelelwe kumaziko amakhulu edatha alawulwa liqela elincinci leenkampani.
Ngokunxuseneyo, uMistral uqalisile ukusebenza namaqabane aphesheya kwindawo eyaziwa ngokuba yi I-AI yoMzimbaPhakathi kwentsebenziswano ekhankanywe yi-Singapore ye-HTX yesayensi kunye ne-arhente yezobuchwepheshe kwiirobhothi, i-cybersecurity, kunye neenkqubo zokukhusela umlilo; kunye nesiJamani Ukunceda, igxininise ekukhuseleni, kunye neemodeli zesenzo solwimi lombono kwiidrones; kunye nabavelisi beemoto abafuna Abancedisi be-AI kwigumbi lokuhlala ukusebenza kakuhle kwaye kuyalawuleka.
Impembelelo eYurophu: ulawulo lwedijithali kunye nenkqubo yendalo yoluntu-yabucala
Ngaphandle kwemiba yobugcisa, iMistral iye yaba luphawu kwingxoxo-mpikiswano Ulongamo lwedijithali eYurophuNangona inkampani izichaza ngokuthi "intsebenziswano ye-transatlantic" -kunye namaqela kunye noqeqesho lomzekelo olusasazeka phakathi kweYurophu kunye ne-United States-, ukuzibophelela kwayo ekuvuleni iimodeli ngenkxaso eqinile kwiilwimi zaseYurophu kuye kwamkelwa kakuhle ngamaziko oluntu kwilizwekazi.
Inkampani iye yavala izivumelwano umkhosi wamaFrentshi, iarhente yengqesho yoluntu yaseFransi, urhulumente waseLuxembourg, kunye neminye imibutho yaseYurophu unomdla wokuhambisa i-AI phantsi kwezicwangciso ezingqongqo zolawulo kunye nokugcina ulawulo lwedatha ngaphakathi kwe-EU. Ngokunxuseneyo, iKhomishini yaseYurophu iye yabonisa a iqhinga lokunyusa izixhobo ze-AI zaseYurophu ezomeleza ukhuphiswano lwemizi-mveliso ngaphandle kokuncama ukhuseleko nokomelela.
Imeko ye-geopolitical nayo ityhala ummandla ukuba usabele. Kuyaqondwa ukuba IYurophu iwile emva kwe-United States ne-China Kugqatso lweemodeli zesizukulwana esilandelayo, ngelixa kumazwe afana ne-China evulekileyo ezinye iindlela ezifana ne-DeepSeek, i-Alibaba, kunye ne-Kimi zivela kwaye ziqala ukukhuphisana nezisombululo ezifana ne-ChatGPT kwimisebenzi ethile, i-Mistral izama ukugcwalisa inxalenye yeso sikhewu ngeemodeli ezivulekileyo, eziguquguqukayo ezihambelana neemfuno zokulawula zaseYurophu.
Ngokwezemali, ukuqaliswa kunyukile 2.700 iidola ezigidi kwaye uye wangena kuqingqo-maxabiso kufutshane 14.000 izigidiLa manani asezantsi kakhulu kunalawo ezikhulu ezifana ne-OpenAI okanye i-Anthropic, kodwa abalulekile kwi-ecosystem yaseYurophu. Inxalenye enkulu yemodeli yeshishini ibandakanya ukunikela, ngaphaya kobunzima obuvulekileyo, iinkonzo zokwenza ngokwezifiso, izixhobo zokusasaza, kunye neemveliso zoshishino ezifana neMistral Agents API okanye i-Le Chat suite enokudityaniswa kwenkampani.
Ukuma kucacile: ukuba a umboneleli weziseko ezingundoqo ze-AI ezivulekileyo neziguquguqukayo evumela iinkampani zaseYurophu (kunye nezinye zengingqi) ukuba zivelise izinto ezintsha ngaphandle kokuxhomekeka ngokupheleleyo kumaqonga ase-US, ngelixa zigcina ulawulo oluthile malunga nokuba iimodeli ziqhutywa phi kwaye njani, kunye nokuququzelela ukuhlanganiswa kunye nezixhobo esele ziphunyeziwe kwiinkqubo zabo.
Ingxoxo malunga nokuvuleleka kokwenyani kunye nemingeni elindileyo
Ngaphandle komdla owenziwe yiMistral 3 kwinxalenye yetekhnoloji yoluntu, akukho kunqongophala kwamazwi abalulekileyo abuza umbuzo. ukuba le mifuziselo inokuqwalaselwa kangakanani na ngokwenene "Vula Umnikezi"Inkampani iye yakhetha indlela ubunzima obuvulekileyoIkhupha ubunzima bokusetyenziswa kunye nokulungelelaniswa, kodwa kungekhona zonke iinkcukacha malunga nedatha yoqeqesho kunye neenkqubo zangaphakathi ezifunekayo ukuvelisa imodeli ukusuka ekuqaleni.
Abaphandi abanje Andreas Liesenfeld, umseki we-European Open Source AI Index, Babonisa ukuba i-bottleneck enkulu ye-AI eYurophu ayikho nje ukufikelela kwiimodeli, kodwa ku idatha yoqeqesho olukhuluUkusuka kuloo mbono, iMistral 3 inegalelo ukuphucula uluhlu lweemodeli ezisebenzisekayoNangona kunjalo, ayisombululi ngokupheleleyo ingxaki esisiseko ye-ecosystem yaseYurophu eqhubeka nokusokola ukuvelisa nokwabelana ngedatha ekumgangatho ophezulu.
I-Mistral ngokwayo iyavuma ukuba iimodeli zayo zesicwangciso esivulekileyo "zisemva kancinane" kwizisombululo ezivaliweyo, kodwa Ugxininisa ukuba umsantsa uyancipha ngokukhawuleza. kwaye ingongoma ephambili kukuba umlinganiselo weendleko nenzuzoUkuba imodeli engenamandla kancinane inokubekwa ngexabiso eliphantsi, ilungiswe kakuhle ukwenzela umsebenzi othile, kwaye ibaleke kufutshane nomsebenzisi, Oku kunokuba nomdla kwiinkampani ezininzi kunemodeli ephezulu enokufikelelwa kuphela nge-API ekude.
Nangona kunjalo, imingeni isekhona: ukusuka kwi- ukhuphiswano oluqatha lwamazwe ngamazwe Oku kunabela kwimfuno yokuqinisekisa ukhuseleko, ukulandeleka, kunye nokuthotyelwa kwemithetho kwiimeko ezifana nokhathalelo lwempilo, imali, kunye norhulumente. Ibhalansi phakathi kokuvuleka, ukulawula, kunye noxanduva luya kuqhubeka nokukhokela iMistral kunye nabanye abadlali baseYurophu kwiminyaka ezayo.
Ukuphehlelelwa IMistral 3 Iyomeleza imbono yokuba i-AI yokusika akufuneki iphelele kwiimodeli ezinkulu, ezivaliweyo.kwaye ibonelela ngeYurophu - kunye nawo nawuphi na umbutho oxabisa ulongamo lwetekhnoloji - ipalethi yezixhobo ezivulekileyo ezidibanisa imodeli yomda we-multimodal kunye noluhlu lweemodeli ezikhaphukhaphu ezikwaziyo ukusebenza emphethweni, ngaphandle kweintanethi, kunye nenqanaba lokwenza ngokwezifiso kunzima ukuhambelana ngamaqonga abucala.
Ndingumntu othanda itekhnoloji ojike umdla wakhe we "geek" waba ngumsebenzi. Ndichithe ngaphezulu kweminyaka eli-10 yobomi bam ndisebenzisa itekhnoloji yokusika kwaye ndikhenkceza ngazo zonke iintlobo zeenkqubo ngenxa yokufuna ukwazi okumsulwa. Ngoku ndiqeqeshelwe ubugcisa bekhompyutha nakwimidlalo yevidiyo. Oku kungenxa yokuba ngaphezu kweminyaka emi-5 ndibhalela iiwebhusayithi ezahlukeneyo kwitekhnoloji kunye nemidlalo yevidiyo, ndisenza amanqaku afuna ukukunika ulwazi oludingayo ngolwimi oluqondakalayo kuye wonke umntu.
Ukuba unayo nayiphi na imibuzo, ulwazi lwam lusuka kuyo yonke into enxulumene nenkqubo yokusebenza yeWindows kunye ne-Android yeefowuni eziphathwayo. Kwaye ukuzinikela kwam kukuwe, ndihlala ndikulungele ukuchitha imizuzu embalwa kwaye ndikuncede usombulule nayiphi na imibuzo onokuba nayo kweli lizwe le-intanethi.