- IMantshi, evela esiqalisweni se-French Mistral AI, iyimodeli yokuqala ye-AI yaseYurophu egxile ekucabangeni kwesinyathelo ngesinyathelo, ngezilimi eziningi.
- Ikhishwe ngezinguqulo ezimbili: Encane (umthombo ovulekile, amapharamitha ayizigidi eziyizinkulungwane ezingama-24.000) kanye Nemaphakathi (yamabhizinisi), enokusebenza okuphezulu kumabhentshimakhi wamazwe ngamazwe.
- IMantshi igqama ngokubeka izinto obala, ukufunda nokusekela izilimi eziningi zaseYurophu, okuyenza ihluke kumamodeli aseMelika naseShayina.
- Imodeli ifuna ukubeka i-Europe phambili ku-AI yomhlaba wonke, ithuthukisa ubukhosi bezobuchwepheshe kanye nokuvuleleka ngamalayisense omphakathi.
Ezinsukwini zamuva nje, iYurophu ingene ngokuphelele emncintiswaneni wamazwe ngamazwe wobuhlakani bokwenziwa ngenxa yokwethulwa kweMagistral, imodeli yokuqala ye-AI egxile ekucabangeni okuthuthukile okuthuthukiswe kuleli zwekazi. Ngemuva kwalesi sigameko kukhona inkampani entsha yaseFrance iMistral AI , enesiphakamiso sayo, efuna ukunikeza enye indlela eqinile nevulekile yokulawula i-US kanye namandla akhulayo eShayina kulo mkhakha wezobuchwepheshe.
Kuze kube manje, ukusungula izinto ezintsha kumamodeli e-AI anamakhono okucabanga ngesinyathelo ngesinyathelo bekuholwa yizinkampani ezifana ne-OpenAI, i-Google, kanye ne-DeepSeek eShayina , okushiya iYurophu isesimweni esicacile sokuthembela kwezobuchwepheshe. Nge-Magistral, i-Mistral AI izibeka njengezwi laseYurophu kulo mkhakha weqhinga, ithola ukwesekwa okuvela kuhulumeni waseFrance kanye nezimali ezinkulu zokutshalwa kwezimali, futhi ibiza cishe amaRandi ayizigidi eziyizinkulungwane ezingu-6.200.
Isimemezelo sikaMagistral senziwe ngesikhathi seLondon Tech Week, enye yemibukiso yobuchwepheshe ehamba phambili emhlabeni wonke, futhi simelela ingqophamlando yale nkampani entsha, kanye nokuthumela umlayezo ocacile mayelana namakhono aseYurophu okuthuthukisa ubuhlakani bokwenziwa besizukulwane esilandelayo ngezindinganiso zokusobala, ukubambisana kanye nobumfihlo.
Yini eyenza iMantshi ihluke kwezinye izinhlobo?
Isici esiyinhloko seMagistral ukugxila kwayo ekucabangeni okucacile , okwenza ikwazi ukuxazulula izinkinga eziyinkimbinkimbi ngenqubo yokucabanga enengqondo, efana nendlela abantu abacabanga ngayo. Ngokungafani nama-LLM endabuko athembele kuphela ekubikezelweni kombhalo, le modeli entsha ingalandela imithetho, yenze izinqumo ezihlelekile, futhi ithethelele nezimpendulo zayo isinyathelo ngesinyathelo.
Kunezinhlobo ezimbili: iMagistral Small , enamapharamitha ayizigidigidi ezingama-24.000 futhi ikhishwe ngaphansi kwelayisensi yomthombo ovulekile ku-Hugging Face; kanye neMagistral Medium , eqondiswe emkhakheni webhizinisi onamandla amakhulu okusebenzisa amakhompyutha futhi etholakala ngama-API namapulatifomu amafu njenge-IBM WatsonX, i-Azure AI, noma i-Amazon SageMaker. Ngenxa yalokhu kuguquguquka, ingafakwa kokubili ekusetshenzisweni kwendawo kanye nezinsizakalo zamafu ezithuthukisiwe.
Ngokusebenza kahle , iMagistral Medium ithole ukunemba okungu-73,6% ku-AIME2024 yamazwe ngamazwe (kanye no-90% ngokusebenzisa ukuvota okuningi), kanti uSmall uthole u-70,7%. Le miphumela ibeka imodeli yaseYurophu ezingeni elifanayo ne-Chinese DeepSeek R1 enamandla futhi isondele kakhulu ku-benchmarks ehamba phambili emhlabeni wonke ekhishwe yi-OpenAI ne-Google, yize isahluleka ukuhambisana nezinguqulo zayo zakamuva.
Ukuqamba kabusha okuvulekile nezilimi eziningi
Amandla eMagistral afaka phakathi ukuvuleka kwayo : ngokungafani nezinqubomgomo ezivimbelayo ze-OpenAI noma i-Google, iMagistral ivumela umphakathi ukufinyelela ezisindweni zemodeli encane, okwenza kube lula ngabathuthukisi, izikhungo, kanye nezinhlangano zomphakathi ukuhlola, ukuzivumelanisa, ngisho nokwenza ngezifiso i-AI ngokwezidingo zabo. Lokhu kubaluleke kakhulu emikhakheni elawulwayo noma ezindaweni ezidinga ukucaca kwezobuchwepheshe.
Ngaphezu kwalokho, iMagistral iqeqeshwe ukucabanga nokukhiqiza izimpendulo ngqo ezilimini eziningi zaseYurophu , njengesiFulentshi, isiSpanishi, isiJalimane, nesiNtaliyane, ngaphandle kokuthembela esiNgisini njengolimi oluphakathi. Lokhu kuqeda isithiyo solimi lwamamodeli endabuko futhi kuyenza ithuluzi eliguquguqukayo kakhulu lokusetshenziswa kulo lonke i-European Union, liletha imiphumela yendabuko nenembile ezinhlamvwini ezahlukene zezinhlamvu kanye nezimo zolimi.
Ngakolunye uhlangothi, inkampani igcizelela ukuhlolwa kanye nokulandelelana kwenqubo yokucabanga: isinqumo ngasinye singalandelwa ngokulandelana, isici esibalulekile emikhakheni efana nomthetho, ubunjiniyela, ezempilo noma ukuphathwa komphakathi.
Izinhlelo zokusebenza nomongo wokuncintisana
Inkampani yaseFrance iklame iMantshi ngezinhlelo zokusebenza ezidinga ukucabanga okuthuthukile engqondweni: ukuxilonga okuhlelekile, ukucabanga okuhleliwe, izihlahla zezinqumo, nezinhlelo ezisekelwe emithethweni. Ngenxa yemodi yayo ye-Flash Response, imodeli ingaletha imiphumela esheshayo ephindwe ka-10 emisebenzini ethile isebenzisa i-chatbot yayo, i-Le Chat, emele enye indlela yase-Europe ku-ChatGPT.
Emkhakheni wamazwe ngamazwe, iMagistral ijoyina impi yamamodeli okucabanga okuthiwa anengqondo , ekuqaleni eholwa yi-OpenAI ngo-2024 futhi ngokushesha yalandelwa yi-DeepSeek eShayina. Nakuba ingakafaki amakhono okubona noma okulalelwayo amaningi njengezimbangi zayo ezinamandla kakhulu, indlela yayo yomthombo ovulekile yenza kube inketho ekhangayo kakhulu kubathuthukisi kanye nezinkampani zomphakathi noma ezizimele ezifuna ukuzimela kwezobuchwepheshe.
Ngemuva kwale phrojekthi kunonjiniyela bangaphambili be-Meta ne-Google, abafana no-Arthur Mensch kanye no-Guillaume Lample, abaletha irekhodi eliqinisekisiwe lesipiliyoni ekwakhiweni kobuhlakani bokwenziwa obusezingeni eliphezulu.
Ukuvela kweMagistral kuphawula isigaba esisha sobukhosi bezobuchwepheshe baseYurophu ku-AI : ukuzibophezela ekubonakaleni, ukuvuleleka, kanye nokuzivumelanisa nezimo zezilimi namasiko zezwekazi, okunikeza okunye okucacile kuneziqhwaga zaseMelika nase-Asia. Nakuba kusenendlela okufanele ilandelwe ukuze kulinganiswe abaholi bomhlaba maqondana namandla aphelele, lokhu kuthuthuka kubeka iYurophu phambili ekuthuthukiseni amamodeli okucabanga nge-AI , kuvula amathuba amasha ezinkampani, izinhlangano zomphakathi, kanye nezakhamuzi ezifuna izixazululo zobuchwepheshe ezifinyeleleka kalula nezilawulekayo.
