Amamodeli omthombo ovulekile
Ithegi yethuluzi le-AIEZZ
Ubheka imikhiqizo ye-AI enethegi ethi “Amamodeli omthombo ovulekile”, enemiphumela efanayo engu-8.
LongCatIzingxoxo ze-AII-LongCat iyimodeli yolimi enkulu yomthombo ovulekile eyethulwe yi-Meituan. Isebenzisa ukwakheka kochwepheshe abaxubile okunamapharamitha ayizigidi eziyizinkulungwane ezingu-560, futhi ngesikhathi sokusebenza ivula ngamandla cishe amapharamitha ayizigidi eziyizinkulungwane ezingu-27. Iklanyelwe ukuqonda umbhalo, ukucabanga okuyinkimbinkimbi kanye nemisebenzi yezingxoxo. Abasebenzisi bangayinika amadokhumenti amade noma izingxoxo eziqhubekayo ukuze izicubungule, okuyenza ifanele abathuthukisi, abasebenzisi bamabhizinisi kanye nabacwaningi bobuhlakani bokwenziwa abahlola ukuhlaziywa kwamadokhumenti, usizo lokubhala amakhodi, ukucabanga kwezibalo kanye nezinqubo zama-ejenti ahlakaniphile. I-LongCat isekela okokufaka okufika kuzinhlamvu ezingu-128k futhi inikeza izindlela zokuyisebenzisa mahhala. Qaphela ukuthi imiphumela yangempela, isivinini sokuphendula kanye nezindleko zokuyisebenzisa zisazothonywa uhlobo lomsebenzi, indawo esebenza kuyo kanye nendlela ethile yokuyibiza.138.5
JanusAmathuluzi okudweba nge-AII-Janus iyithuluzi le-AI elinemodi eziningi lokudweba elenzelwe ukudala izithombe nokuqonda okuqukethwe okubukwayo. Ulwazi lukhombisa ukuthi okwamanje iyasebenza futhi inikeza indlela yokuyisebenzisa mahhala. Abasebenzisi bangakha izithombe besebenzisa iziqondiso zombhalo, basebenzise namandla ayo okuqonda izithombe ukuhlaziya izigcawu, bakhiphe ulwazi kumagrafu ezemfundo, futhi baguqule amafomula ezibalo abe ikhodi ye-LaTeX. Ilungele abadali bokuqukethwe okubukwayo, abacwaningi, onjiniyela, abahleli bokuqukethwe kwabezindaba abazimele kanye nabaqeqeshi bezemfundo, ukuze benze imidwebo yokuqala yemibono, izithombe ezihambisana nokuqukethwe, incazelo yamagrafu kanye nokuhlela amafomula. Ekusetshenzisweni kwangempela, imininingwane yezithombe ezikhiqiziwe, imiphumela yokubonwa kwamagrafu kanye nokunemba kwekhodi kusafanele kuqinisekiswe ngumsebenzisi.077.6
DiffRhythmAmathuluzi omsindo e-AII-DiffRhythm iyimodeli ye-AI yokukhiqiza izingoma eyakhiwe yi-ASLP Laboratory yaseNorthwestern Polytechnical University. Isebenzisa imodeli yokusabalalisa efihlekile, futhi ihloselwe abadali nabathuthukisi abadinga ukwenza ngokushesha izibonelo zezingoma, umculo wangemuva noma ucwaningo lwamamodeli. Ngemva kokufaka amagama engoma nomyalo wesitayela, ithuluzi lingakhiqiza ingoma ephelele enezwi lomculi kanye nokuphelezelwa komculo. Isikhathi esiphezulu sokukhiqiza ngesikhathi esisodwa cishe siyimizuzu engu-4 nemizuzwana engu-45, futhi lisekela izingoma zesiShayina nesiNgisi. Le phrojekthi inikeza amamodeli avulekile nekhodi yokwenza inference, ngakho ilungele ukuzama, ucwaningo kanye nokuthuthukiswa okuqhubekayo. Imiphumela ekhiqiziwe isengadinga ukukhethwa, ukulungiswa noma ukukhiqizwa kwangemva komuntu, futhi umphumela wangempela uthonywa amagama engoma nemiyalo yesitayela.018.5
K2 ThinkIzingxoxo ze-AII-K2 Think iyimodeli yokucabanga yomthombo ovulekile eyethulwe ngokuhlanganyela yi-Mohamed bin Zayed University of Artificial Intelligence yase-United Arab Emirates kanye ne-G42. Igxile ekusingatheni umqondo onzima, izinkinga zezibalo kanye nemibuzo yezindawo ezikhethekile. Imodeli inamapharamitha ayizigidi eziyizinkulungwane ezingama-32, futhi isekela ukukhiqizwa kwekhodi kanye nemibuzo yolwazi emikhakheni efana nefiziksi nekhemistri. Abasebenzisi bangathumela imibuzo ngesixhumi esibonakalayo sengxoxo, baqhubeke nokubuza ngokusekelwe ezimpendulweni, okwenza ilungele onjiniyela, abacwaningi, othisha kanye nabasebenzisi abangochwepheshe abadinga izingxoxo ezigxile ekuhlaziyeni. Ulwazi olutholakalayo lubonisa ukuthi ingafinyelela isivinini sokucabanga esingaphezu kwama-token angu-2000 ngomzuzwana, futhi igcizelela ukusebenza kahle okuphezulu kwamapharamitha. Izimpendulo zangempela zisadinga ukuqinisekiswa ngumuntu, ikakhulukazi ezimweni ezidinga ukunemba okuphezulu njengezibalo, ucwaningo kanye nezezimali.038.3
Liquid AIIbhokisi lamathuluzi e-AII-Liquid AI iyibhokisi lamathuluzi e-AI elenzelwe amadivayisi asemaphethelweni, umkhiqizo wayo oyinhloko i-Liquid Foundation Models (LFMs), enosayizi abasuka kumapharamitha angu-350M kuye ku-2.6B futhi esekela izinhlobo zamamodeli ombhalo, olimi nombono kanye nomsindo. Abathuthukisi bangasebenzisa inkundla ye-LEAP ukulanda amamodeli, ukuwenza kahle nokuwasebenzisa kumadivayisi, noma basebenzise uhlelo lweselula i-Apollo ukuhlola endaweni. Le nkundla ihloselwe abathuthukisi abazimele, amabhizinisi asaqalayo, abacwaningi, abakhiqizi bezingxenyekazi zekhompuyutha kanye namabhizinisi adinga izixazululo ezenziwe ngokwezifiso, futhi ifanele amatheminali afana namafoni, amakhompyutha aphathekayo, izimoto namarobhothi. Amamodeli atholakala ku-Hugging Face; imiphumela ethile isancike ezimweni zehadiwe, kusayizi wemodeli nasekuhambisaneni kokusetshenziswa.057.8
ModelScopeIbhokisi lamathuluzi e-AII-ModelScope iyinkundla yezinsizakalo zamamodeli yomthombo ovulekile eyethulwe yi-DAMO Academy yakwa-Alibaba. Ihloselwe abathuthukisi be-AI, abacwaningi, amaqembu ezinkampani nezikhungo zemfundo, futhi inikeza uchungechunge lwamathuluzi olusuka ekuhloleni amamodeli luye ekwenzeni inference, ekuqeqesheni, ekusabalaliseni nasekuthuthukiseni izinhlelo zokusebenza. Le nkundla ihlanganisa amamodeli angaphezu kuka-300, ahlanganisa umbono wekhompyutha, ukucutshungulwa kolimi lwemvelo, inkulumo, ama-multimodal kanye ne-computing yesayensi, futhi inikeza namandla okwabelana ngamamodeli, ukuphatha ama-dataset kanye nokusabalalisa efwini. Abasebenzisi bangathola amamodeli ngokomsebenzi, bawahlole, bawenze i-fine-tuning noma bawaqeqeshe, bese bewaxhuma kumaphrojekthi besebenzisa umtapo wePython nezixhumanisi ezihlanganisiwe. Ekusetshenzisweni kwangempela, izindleko zokuhambisana kusafanele zihlolwe ngokusekelwe kokuncika kwemodeli, kumandla ekhompyutha nasendaweni yokusabalalisa.058.6
WhisperAmathuluzi omsindo e-AII-Whisper iwuhlelo lokubona inkulumo ngokuzenzakalelayo olwakhiwe yi-OpenAI, olugxile ekuguquleni umsindo ube umbhalo, ekuhumusheni inkulumo nasekuhleleni okuqukethwe. Abasebenzisi bangafaka umsindo ohlelweni ukuze bathole umbhalo ohambisanayo, futhi babuke izitembu zesikhathi ngokwegama noma ngomusho; kwezinye izinkulumo ezingezona ezesiNgisi, ingawuhumushela ngokuqondile umbhalo wesiNgisi. Yaqeqeshwa ngamahora angu-680,000 edatha egadiwe ehlanganisa izilimi nemisebenzi eminingi. Imodeli nekhodi kuyatholakala ngokukhululekile, ngakho ilungele ukuhlanganiswa konjiniyela, ukuhlolwa ngabacwaningi, kanye nokuhlela izingxoxo, amaphodikasti, imihlangano, izifundo nokuqukethwe kwamavidiyo. Imiphumela yokubona inkulumo kusafanele ihlolwe umuntu ngokucabangela ikhwalithi yokuqoshwa, indlela yokuphimisa namagama ochwepheshe.028.4
Stable DiffusionUkudala imidwebo nge-AII-Stable Diffusion iwuchungechunge lwamamodeli e-Stability AI okukhiqiza izithombe. Ikhasi elisemthethweni lamanje lamamodeli ezithombe ligxile ku-Stable Diffusion 3.5, ehlinzeka nge-Large, Turbo ne-Medium: i-Large igxile kukhwalithi yezinga lochwepheshe nasekulandeleni imiyalelo, i-Turbo igcizelela ukukhiqiza okusheshayo ngezinyathelo ezimbalwa, kanti i-Medium ibhalansisa ikhwalithi, ukwenza ngokwezifiso nokusebenza ku-hardware yabathengi. Amamodeli angasetshenziswa ngelayisensi yokuzisingatha, nge-Stability AI API, ngabalingani bamafu noma ngezinhlelo zokusebenza zewebhu, futhi ahambisana nezinsizakalo zokusula, ukubuyisela izingxenye zesithombe, ukunweba isithombe, ukukhulisa usayizi, kanye nokulawula okusekelwe kumidwebo, esakhiweni nasesitayeleni. Ikheli le-GitHub repository elashicilelwa ekuqaleni manje libuyisa u-404; ukusetshenziswa nokusetshenziswa kwezohwebo kufanele kuqondiswe emakhasini asemthethweni amamodeli ezithombe namalayisensi.049Wonke amathuluzi analeli thegi ayaboniswa.