Longcat-2.5-preview Uyasebenza

I-LongCat-2.5-Preview iyimodeli yolimi enkulu yesizukulwane esisha eyethulwe yi-Meituan. Igcina amapharamitha aphelele angu-1.6T e-LongCat-2.0, cishe amapharamitha asebenzayo angu-48B kanye nomongo wamathokheni angu-1M, futhi yengeza amandla emvelo e-multimodal, ukuze yenze imisebenzi emide ezindaweni ezifana namatheminali, iziphequluli kanye nesofthiwe yedeskithophu. I-LongCat-2.5-Preview isivele ixhunywe ku-API esemthethweni, iyahambisana nezindawo zokuxhumana ze-OpenAI ne-Anthropic, ikhiphe amathokheni afinyelela ku-128K, futhi isekela amathuluzi e-Agent adumile afana ne-Codex ne-OpenClaw ukuthi ayibize ngokuqondile.

Ukuqonda i-multimodal: Yengeza amandla okuqonda izithombe, ihlaziye okuqukethwe kwezithombe, futhi isekele imibuzo nezimpendulo ezinqamula izindlela, ukufingqa okuqukethwe kanye nokucabanga okubukwayo okuyinkimbinkimbi.
Imisebenzi emide ye-Agent: Yenza imisebenzi emide ezindaweni ezifana namatheminali, iziphequluli, i-GUI, amaspredishithi namathuluzi okuklama.
Amandla e-Coding: Isebenza kahle ekukhiqizeni ikhodi, ekuqondeni ikhodi nasemisebenzini yokuhlela ezenzakalelayo, iqhubeke nokugxila kwe-LongCat-2.0 ku-Agentic Coding.
Ukuhambisana okujulile nezindawo zokuthuthukisa: Isebenzisana kahle namathuluzi adumile afana ne-Claude Code, Hermes, OpenClaw, OpenCode ne-Kilo Code.
Izindawo ezimbili zokuxhumana: I-API iyahambisana ngasikhathi sinye namafomethi e-OpenAI ne-Anthropic, ikhiphe amathokheni afinyelela ku-128K, futhi amathuluzi e-Agent akhona angaxhumeka ngokuqondile.

Isakhiwo se-MoE esisebenzisa amapharamitha ngokukhetha: Iqhubeka nendlela ye-Mixture of Experts, enamapharamitha aphelele angu-1.6 trillion, kodwa ekucabangeni ngakunye isebenzisa cishe amapharamitha ayizigidi eziyizinkulungwane ezingu-48 kuphela. Ngokusebenzisa umzila oguqukayo, ibiza iqoqo lochwepheshe elihambisana kakhulu nomsebenzi wamanje, ngaleyo ndlela igcine umthamo wemodeli futhi ilawule izindleko zokubala.
I-multimodal yemvelo ihlanganiswe esisekelweni: Ngokungafani nenkathi ye-2.0, lapho izithombe namavidiyo zazihlukaniswa zibe amakhosombe azimele avulekile, i-2.5 ihlanganisa amandla e-multimodal ngqo emzimbeni oyisisekelo, ukuze ulwazi olubukwayo nombhalo kumodelwe ngokuhlanganyela ngaphakathi kwemodeli efanayo. Le ndlela isuselwa ku-DiNA (Discrete Native Autoregressive) paradigm eqinisekiswe yi-LongCat-Next, eguqula izindlela zibe amathokheni ahlukene ukuze zimodelwe ndawonye.
Ukusekelwa komongo omude kakhulu: Iwindi lomongo wemvelo lamathokheni angu-1M lingamukela amadokhumenti amade kakhulu, ikhosombe lonke lekhodi, amalogi esistimu kanye nemisebenzi eyinkimbinkimbi enemijikelezo eminingi ngesikhathi esisodwa. Lokhu kunikeza isisekelo senkumbulo sokwenziwa kwemisebenzi eminingi ye-Agent yesikhathi eside, futhi kuwumphumela wobunjiniyela bezinhlelo osekelwa yizinqwaba zamakhompyutha ase-China ezikwazi ukuphatha amapharamitha amakhulu nomongo wezinga lesigidi ngesikhathi esisodwa.
Ukugxila ekuqeqeshweni kwe-Agentic: Imodeli yaqeqeshwa kusukela ekwakhiweni kwayo igxile ezimweni ze-Agent, isekela ukubizwa kwamathuluzi nokucabanga ngezinyathelo eziningi ngokwemvelo, futhi yehlisa ukubambezeleka kwemisebenzi enamaketanga amade ngobuchwepheshe bokusheshisa ukucabanga obufana ne-dlp speculative decoding.

Bhalisa i-akhawunti: Vula https://longcat.ai/platform/ ubhalise futhi ungene ku-LongCat Open Platform ye-Meituan.
Thola i-Key: Dala i-API Key kukhonsoli.
Faka i-SDK: Sebenzisa i-OpenAI noma i-Anthropic official SDK ekhona, ngaphandle kokufaka iklayenti elisha.
Lungisa ukumisa: Shintsha i-base_url ibe ngu-https://api.longcat.ai/openai (ifomethi ye-OpenAI) noma https://api.longcat.ai/anthropic (ifomethi ye-Anthropic), bese ufaka i-Key.
Cacisa imodeli: Faka i-LongCat-2.5-Preview njenge-ID yemodeli, bese uthumela izicelo ngefomethi ejwayelekile ye-OpenAI chat.completions noma ye-Anthropic messages.
Setha ANTHROPIC_BASE_URL=https://api.longcat.ai/anthropic kanye ne-ANTHROPIC_AUTH_TOKEN=I-Key yakho eziguquguqukayo zemvelo.
Setha i-ANTHROPIC_MODEL ibe ngu-LongCat-2.5-Preview.
Ngemva kokuqala i-Claude Code, ungabiza le modeli ngokuqondile ukuze yenze imisebenzi yokubhala ikhodi. Kusebenza ngendlela efanayo ku-OpenClaw, OpenCode, Codex nakwamanye amathuluzi: shintsha i-base_url negama lemodeli.
Ngemva kokungena kuwebhusayithi esemthethweni ye-LongCat, khetha imodeli ye-LongCat-2.5-Preview ukuze uxoxe, ulayishe izithombe futhi uzame imisebenzi elula ye-Agent kuwebhu.

Imodeli enkulu esebenzisa amapharamitha ngokukhetha + ukubala okubiza kancane: Ku-1.6T wamapharamitha aphelele, kusebenze cishe angu-48B kuphela, kuhlanganiswe amandla emodeli esezingeni eliphezulu nezindleko zokucabanga ezilawulekayo.
Umongo wemvelo wezinga lesigidi: Iwindi lamathokheni angu-1M lingalayisha ikhosombe lonke lekhodi, amadokhumenti amade noma amalogi aphelele ngesikhathi esisodwa, linikeze imisebenzi emide isisekelo senkumbulo.
I-multimodal yemvelo esisekelweni: Ukuqonda izithombe nombhalo kumodelwa ngokuhlanganyela ngaphakathi kwemodeli efanayo, futhi akusahlukaniswa kube amakhosombe azimele, okwenza ukuxhumana kube ngokuhlangene kakhulu.
Amandla e-GUI Agent emisebenzini emide: Kusukela emisebenzini yekhodi kuya ekusebenzeni kwangempela kwamatheminali, iziphequluli, amaspredishithi namathuluzi okuklama, kuhlanganisa izimo zokusebenza eziphelele.
Ukuxhumeka ngaphandle kokuthutha phakathi kwamaphrothokholi amabili: Iyahambisana nezindawo zokuxhumana ze-OpenAI ne-Anthropic ngesikhathi esisodwa, futhi amathuluzi e-Agent akhona angayibiza ngokuqondile ngokushintsha i-base_url kuphela.

Umsizi wokuhlela osezingeni lekhosombe lonke lekhodi: Sebenzisa umongo wamathokheni angu-1M ukulayisha ikhosombe lonke ngesikhathi esisodwa, wenze ukwakhiwa kabusha kwamafayela amaningi, uthole amaphutha futhi ukhiqize ama-PR amakhulu, uqondiswa ngokuqondile yi-Claude Code noma i-OpenCode.
Ukuzenzakalela kwemisebenzi emide ezinhlelweni ezahlukene: Vumela imodeli ukuthi isebenzise isiphequluli, ispredishithi kanye nesofthiwe yedeskithophu ukuze yenze imisebenzi yezinyathelo eziningi engadingi ukugadwa, efana nokuthi “qoqa idatha yewebhu → yihlanze → uyifake ku-Excel → ukhiqize umbiko”.
Ukuhlaziywa kwamalogi namadokhumenti ezinga lebhizinisi: Faka amalogi esistimu ezinga lamathokheni ayisigidi, amafayela ezinkontileka noma imibhalo yezifundo ngokuphelele, wenze imibuzo nezimpendulo ezijulile, ukufingqa kanye nokutholwa kokungajwayelekile ngesikhathi esisodwa, ugweme ukunqamuka komongo okubangelwa ukusikwa kwe-RAG.
Ukuhlolwa nokugcinwa kwe-GUI: Sebenzisa amatheminali nesofthiwe yedeskithophu ngokuzenzakalela ukuze wenze ukumisa ngobuningi, ukuhlolwa kokubuyela emuva kwesixhumi esibonakalayo kanye nokufakwa kwedatha okuzenzakalelayo ezinhlelweni ezindala, phakathi kweminye imisebenzi edla isikhathi yabantu.
Usizo ekukhiqizeni okuqukethwe kwe-multimodal: Ngokuhlanganisa amandla amasha okuqonda izithombe, yenza ukuguqulwa kwesithombe-skrini kube yikhodi, ubuyisele imiklamo kumathuluzi okuklama, futhi uhlele kabusha noma udale okuqukethwe okusekelwe ezintweni ezihlanganisa izithombe nombhalo.