Mistral Large 4 Uyasebenza

I-Mistral Large 4 iyimodeli ehamba phambili eyakhiwe yi-Mistral AI yase-France, ebizwa ngokuthi «Le Chonk ikati elikhuluphele». I-Mistral Large 4 isebenzisa ukwakheka kwe-MoE okunemininingwane emihle, inamapharamitha angu-1.05T isiyonke, kanti kuvulwa cishe angu-49B kuphela ekusetshenzisweni ngakunye. Iphinde ibe ne-encoder yombono engu-1.6B, isekele i-multimodal ngokwemvelo, futhi ibe nomongo ongu-1M. I-Mistral Large 4 isiyatholakala ku-Mistral API; emiphumeleni yokuhlolwa, ihamba phambili kubunjiniyela besoftware, ezinqubweni zezezimali nasekutholeni indawo ezithombeni zokuzwa kude. Igxile ekuvikelweni kwamanethiwekhi okungafakwa endaweni kanye nasezimweni zokusetshenziswa ezidinga ubukhosi bezinkampani.

Ukuqonda kwemidiya eminingi okwakhelwe ngaphakathi: ikwazi ukufunda umbhalo nezithombe ngesikhathi esisodwa, ikhiphe umbhalo, futhi iwindi lomongo lifinyelela kuma-token ayisigidi esingu-1.
Ubunjiniyela besoftware obunochungechunge olude: ithole u-62% emisebenzini yobunjiniyela yangempela ye-DeepSWE v1.1, ikwazi ukubhala ikhodi nokwenza ukuthuthukiswa okunezinyathelo eziningi.
Ukuhlaziywa kwezimali zamabhizinisi: ithole u-67% esilinganisweni sezinqubo zezimali i-Finch, inekhono lokuthola idatha kumaspredishithi angahlelekile, ukwakha amamodeli nokwenza imibiko.
Ukumaka izindawo ezibonakalayo: ikwazi ukudweba amabhokisi azungeze okuqondiwe ezithombeni eziyinkimbinkimbi, njengezithombe zasemoyeni zokuzwa kude; ithole u-73% ku-DIOR-RSVG, futhi ingasetshenziswa ukuhlola umonakalo wezinhlekelele nokuhlola izintambo zokudlulisa ugesi.
Ukuguqula imidwebo ibe yi-CAD: ingaguqula imidwebo yobuchwepheshe ibe amamodeli e-CAD; ngokusho komkhiqizi, ithole imiphumela ephezulu ezivivinyweni ezihlobene nama-semiconductor.
Ukuvikela amanethiwekhi: inguqulo yokuhlola inikeza izithiyo ezimbalwa nemisebenzi eqinile yokuphepha ku-inthanethi, futhi isetshenziselwa imisebenzi yokuzivikela efana nokuskena ubungozi.
Izilimi eziningi nokusatshalaliswa ngaphansi kobukhosi bedatha: yaqeqeshwa ngezilimi ezingaphezu kuka-160, futhi ingasatshalaliswa kumaseva akho noma ngama-API asezindaweni ezivikelwe umthetho we-European Union.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI anomthombo ovulekile
Isakhiwo se-Mixture of Experts (MoE) esinemininingwane emincane: Imodeli inamapharamitha angu-1.05T esewonke, ahlukaniswe aba amanethiwekhi amancane amaningi “ochwepheshe”. Lapho icubungula ithokheni eyodwa, ivula ingxenye encane kuphela yawo, futhi ifaka i-visual encoder engu-1.6B ecubungula okokufaka kwezithombe. Umgomo uwukushintshanisa ukusebenza okungatheni ukuze kutholakale umthamo wemodeli enkulu kanye nezindleko zokucabanga zemodeli encane; izindleko zokubala zethokheni ngayinye zicishe zilingane nalezo zemodeli eminyene engu-49B.
Ukuqeqeshwa kusukela ekuqaleni namandla okwenza ikhompuyutha angawayo: Imodeli yaqeqeshwa ngokuphelele kusukela ekuqaleni, kwathatha izinyanga ezimbili, kusetshenziswa cishe amakhadi e-NVIDIA Grace Blackwell angu-4000. Konke kwasetshenziswa ezikhungweni zedatha zaseYurophu eziphethwe yi-Mistral. Lokhu kuqinisekisa ukuthi inqubo yokuqeqesha iyalawuleka futhi idatha iyalandelheka, kusekele umbono wayo we-AI ezimele, futhi amakhasimende athole imodeli engenakunqanyulwa uhulumeni wangaphandle.
Ukuqeqeshwa kwangemva kokukhishwa okuqhubekayo kusetshenziswa i-Reinforcement Learning: Okukhishwe yinguqulo yokubuka kuqala, futhi imiphumela ingase ishintshe ngaphambi kokukhishwa kwezisindo. Indlela iwukusebenzisa i-RL ukuqhubeka nokuthuthukisa ukusebenza kwemodeli kumaketanga amade emisebenzi efana nokucabanga, ukubhala ikhodi nezinqubo zezimali.
Ukumiswa kwendawo ebonakalayo (Grounding): Ngaphezu kwamakhono e-multimodal, imodeli ingakhipha izixhumanisi eziqondile zendawo yento esithombeni. Umgomo uwukuthi i-visual encoder ikhiphe izici zesithombe, bese iziqondanisa nemiyalelo yolimi ngokuwela izindlela, ukuze ikwazi ukukhomba into ezingeni lamaphikseli ezimweni ezifana nezithombe zokuzwa kude nemidwebo yobunjiniyela.

Bhalisa i-akhawunti: Vakashela iwebhusayithi esemthethweni ye-Mistral Studio ukuze ubhalise futhi ungene ngemvume.
Dala i-API Key: Khiqiza ukhiye we-API kukhonsoli ukuze uwusebenzisele ukuqinisekisa ukubizwa.
Qinisekisa i-ID yemodeli: Khetha imodeli ethi mistral-large-4.
Khetha indlela yokuxhuma: Ungabiza ngokuqondile i-API https://api.mistral.ai.
Bhala isicelo: Lapho uthumela isicelo, faka i-API Key, igama lemodeli nokuqukethwe kokufakwayo.

Isakhiwo se-Sparse MoE: amapharamitha angu-1.05T esewonke, kodwa kuvulwa angu-49B kuphela ekusebenzeni ngakunye, kuhlanganisa umthamo wemodeli enkulu kakhulu nezindleko zokucabanga ezicishe zifane nezemodeli encane.
Inkampani ithi iyimodeli enesisindo esivulekile enamandla kakhulu eYurophu naseMelika, futhi iyifulegi lokuqala ngemva kokuqoqwa kwezimali kochungechunge D okungama-euro ayizigidi eziyizinkulungwane ezi-3.
Ubukhosi nokulawulwa kokuphepha: ukuqeqeshwa kwenziwa ngokuphelele ezikhungweni zedatha eziphethwe yiYurophu, kanti izisindo zingalandwa futhi zisetshenziswe ngokuzimela, ngaphandle kokuthinteka ukuvinjelwa kokunikezwa okuvela kohulumeni bangaphandle.
Amamodi amaningi omdabu kanye nokuhola ekumakeni indawo ngokubona: umongo wamathokheni ayizigidi ezi-1, nekhono lokukhomba nokudweba amabhokisi ezithombeni elidlula amamodeli avaliwe asezingeni eliphambili.
Imiphumela evelele ekuhlolweni kwangaphambi kokukhishwa: i-DeepSWE ithole u-62% kubunjiniyela besoftware, kanti i-Finch ithole u-67% kwezezimali; womabili ahola izimbangi zawo kuleso sigaba.

Ubunjiniyela besofthiwe obunochungechunge olude: ukuqedela ngokuzimela imisebenzi ephelele yokuthuthukisa, kusukela ekuhlaziyweni kwezidingo kuye ekubhalweni nasekuhlolweni kwekhodi.
Ukuhlaziywa kwezezimali kwamabhizinisi: ukuthola idatha kumaspredishithi namadokhumenti angahlelekile, ukwakha amamodeli nokwenza imibiko, okulungele ukubuyekeza ama-akhawunti nokucwaninga.
Ukuhlolwa kwezithombe zokuzwa kude: izinkampani zomshwalense zihlola umonakalo wezivunguvungu ngezithombe ezithathwe emoyeni, izinkampani zikagesi zihlola izintambo zokudlulisa ugesi, kanti amapulazi aqapha ukukhula kwezitshalo.
Ukuvikela amanethiwekhi: ukuskena ubungozi nokwenza izivivinyo zokuhlasela nokuvikela; ngemva kokulanda ama-weight endaweni, akuvinjwa yizivikelo zamamodeli avaliwe.
Ukuhlolwa komthetho nezinkontileka: ukusiza ukuqedela lonke uhlelo lomsebenzi wezomthetho; izinga lokuphasa lika-Harvey lingu-15%, futhi okwamanje akekho ophumelele kulo lonke ibhentshimakhi, okukhombisa ukuthi la makhono asesigabeni sokuqala.