LongCat-2.5 Uyasebenza

I-LongCat-2.5 iyimodeli yemvelo ye-multimodal eyethulwe yi-Meituan LongCat, enamapharamitha angu-1.6 trillion kanye namapharamitha acishe abe yizigidigidi ezingu-48 asebenzayo ngethokheni ngayinye. I-LongCat-2.5 iqondiswe emisebenzini yama-agent yesikhathi eside, iguqula amasignali okubuka abe amathokheni anezigaba eziningi futhi iwacubungule ngokuhlanganyela nombhalo, iqinise ikhono le-Visual Coding. Ngokusebenzisa i-Agentic RL enkulu, ulayini wokwenziwa kwedatha, kanye nokuqeqeshwa komongo omude nokwenziwa ngcono kwe-W4A4 quantization kumandla ekhompyutha akhiqizwa kuleli, itholakala eqenjini elihamba phambili kanye ne-GPT-5.5 ne-Claude-Opus-5 ekuhlolweni kokuhlela kumadivayisi, ukuphequlula nokusesha, nokuqonda okubukwayo.

Ukuhlela ngama-agent: Iqedela kusukela ekuqaleni kuze kube sekugcineni ukuthuthukiswa kwamarepositri ekhodi ayinkimbinkimbi nokwakhiwa kwamaphrojekthi aphelele.
Ukuqonda okubukwayo kwemvelo: Iphatha izithombe nombhalo ngendlela yamathokheni ahlanganisiwe, futhi inekhono eliqinile ku-Visual Coding.
Ukukhiqizwa kobuciko ne-3D: Isekela ukukhiqizwa kwamakhasi ewebhu aqukethe ukulingisa uketshezi ngesikhathi sangempela, imibukiso yedijithali ye-3D, amamodeli amancane amadolobha, izigcawu ezincane nokunye.
Umsebenzi wasehhovisi ochwepheshe: Uhlanganisa ukuhlaziywa kokusebenza kwebhizinisi, imibiko yempendulo yamakhasimende, ukukhiqizwa kwezimemo zemicimbi nokunye.

Ukumodela izethulo ezibukwayo: Ilandela indlela ye-LongCat-Next, isebenzisa i-residual vector quantization ezicini ze-native resolution ze-encoder ehambisanisa incazelo, imele amasignali okubuka njengamathokheni ahlukene anezigaba eziningi; uma i-tokenizer isilungisiwe, amathokheni okubuka angabalwa kusengaphambili ungaxhunyiwe ku-inthanethi, okwenza kube lula ukupakishwa kokulandelana okude nokulinganiswa komthwalo okusabalalisiwe.
Ukuqiniswa okukhulu kwama-agent: Yakha ulayini wokwenziwa kwedatha onganwetshwa, okhiqiza ngokuzenzakalela imisebenzi, izindawo nezimo zokuhlola eziqinisekisekayo, futhi iqhubeke nokwengeza idatha kusetshenziswa imizila yokwehluleka yama-Agent angempela; isebenzisa ukuqeqeshwa kwe-partial rollout okungavumelani ukunciphisa ukungasebenzi kahle kwemisila emide, ukusikwa kwe-logprob ukucindezela ukuchema kwe-gradient okubangelwa ukuphelelwa yisikhathi kwamasampula, ukunqanyulwa kwenethiwekhi, ukusulwa kwe-cache kanye ne-LLM judge ukuvimbela i-reward hacking, nokuqeqeshwa okuxubile kuyo yonke imisebenzi nama-Harness ukuze kuthuthukiswe ukujwayeleka.
I-Multi-expert online distillation (MOPD): Ochwepheshe bezizinda baqeqesha umfundi besebenzisa i-KL divergence yesilulumagama esigcwele esisampulwe umfundi ngesikhathi sangempela; impendulo kathisha ihlukaniswa yaba ama-per-cut prefix anobude obungaguquki, futhi ubude besampula lomfundi bukhawulelwe ukulawula ukuphambuka kokusatshalaliswa. Ngokuhlanganisa ne-Tree Attention, ukuqeqeshwa kuqedwa ngokudlula okukodwa kokuya phambili nokubuyayo, kuncishiswe izindleko zekhompyutha nokugcina.
Ukuqeqeshwa komongo omude kumandla ekhompyutha akhiqizwa kuleli: Izinga lokufana le-FSDP/EP lihambisana nenani lamakhadi ama-supernode ukuze ukuxhumana kuhlale ngaphakathi kwesizinda somkhawulokudonsa ophezulu, kanti ukufana kwesilulumagama esixubile kuqhubeka kwehlisa ukusetshenziswa kwememori; i-FSDP yokuhlukanisa kabili ihlukanisa izisindo eziyinhloko ze-FP32 nezimo ze-optimizer kuzo zonke izizinda, ukuxhumana okungenakho ukukopisha kuqondaniswe nokubala, futhi i-throughput ye-SFT ekhadini elilodwa inyuka ngo-86% uma iqhathaniswa ne-DDP.
Ukuvumelana kokuqeqesha nokusebenzisa: Isebenzisa ama-operator e-batch-invariant ahlanganisiwe, inweba ukwesekwa kwe-TP ne-DCP, futhi iqinisekisa ukuvumelana kwe-bitwise kwe-logprob yethokheni ngayinye phakathi kwe-Training, Prefill ne-Decode. Ukwenziwa ngcono kwe-Rollout yomongo omude: Ngaphansi kokuhlukaniswa kwe-PD, i-QCP/KVP ne-HiCache kwandisa i-cache esebenzayo, kanti ukusetshenziswa kwenkomba yezendlalelo ye-LSA ekwabelaneni kwenza i-KV prefetch phakathi kwezendlalelo, okusondeza i-Decode cache ekuphindaphindekeni kabili; ukuhlela okuqaphela umthamo kugwema ukuqalisa ngokushintshana kwemizila emide okubangela ukuqhubeka kokubala kabusha.

Ukuhlola ngokushesha: Vakashela ikhasi lokuhlola eliku-inthanethi ku-https://longcat.chat ukuze ungene kusixhumi esibonakalayo sengxoxo, bese ucela okudingayo ngolimi lwemvelo.
Ukuxhuma nge-API: Ngena ku-https://longcat.chat/platform/product ukuze ufake isicelo sokhiye we-API futhi uxhume imodeli kuhlelo lwakho lokusebenza.

Ukusebenza okuhamba phambili: Amapharamitha ayizigidi eziyizigidi ezingu-1.6 kanye namapharamitha ayizigidigidi ezingu-48 asebenzayo ngethokheni ngayinye; iseqenjini elihamba phambili kanye ne-GPT-5.5 ne-Claude-Opus-5 ekuhlolweni okuningi.
I-multimodal yemvelo: Amasignali okubuka namathokheni ahlukene anezigaba eziningi kumodelwa ndawonye nombhalo, futhi ikhono le-Visual Coding livelele.
Ama-agent esikhathi eside: Ngokusekelwa yi-Agentic RL enkulu, iqedela kusukela ekuqaleni kuze kube sekugcineni ukuthuthukiswa kwamarepositri ekhodi ayinkimbinkimbi nokwakhiwa kwamaphrojekthi aphelele.
Ukuthuthuka okusheshayo: Ukuqeqeshwa kokulandelana okude kanye nokuguqula imizila yokwehluleka ibe yimisebenzi yokuqeqesha kunikeza ukuthuthuka okukhulu uma kuqhathaniswa ne-LongCat-2.0.
Ukuhambisana namandla ekhompyutha akhiqizwa kuleli: Ukuthuthukiswa kokufana kwama-supernode ne-W4A4 quantization kwenza i-MoE yezinga le-1.6T iqeqesheke futhi isetshenziswe kahle kuma-accelerator e-AI akhiqizwa kuleli.
Ubunjiniyela obuqinile: Ukuvumelana kwe-bitwise phakathi kokuqeqesha nokusebenzisa kanye nokwenziwa ngcono kwe-rollout yomongo omude kusekela ukusebenza okuzinzile nokuphumelelayo kwe-RL yesikhathi eside ngezinga elikhulu.

Iwebhusayithi esemthethweni yephrojekthi: https://longcat.ai/blog/longcat-2.5/

Ukuthuthukiswa kwesofthiwe: Iqedela kusukela ekuqaleni kuze kube sekugcineni ukuthuthukiswa kwamarepositri ekhodi ayinkimbinkimbi, ukwakhiwa kwamaphrojekthi aphelele nemisebenzi yokuhlela kumadivayisi.
Ukukhiqizwa kwamakhasi ewebhu okubukwayo: Ikhiqiza amakhasi ewebhu ashukumisayo anezinga eliphezulu lamafreyimu aqukethe ukulingisa uketshezi ngesikhathi sangempela, imibukiso esebenzisanayo yezinto zobuciko ze-3D, amamodeli amancane amadolobha namakhasi anemibala yesitayela samaShayina.
Ukuhlaziywa komsebenzi wasehhovisi ochwepheshe: Ihlanganisa idatha yokusebenza evela eziteshini eziningi, ikhiqize ukubalwa kwenzuzo, ukuhlaziywa kwezinto eziqhuba ukusebenza neziphakamiso zokumaketha nezitoko.
Ukucubungula impendulo yamakhasimende: Ikhomba impendulo ephindaphindiwe, iyihlele ngokobucayi, futhi ikhiqize umbiko wokuhlaziya osebenzisanayo we-HTML.
Amathuluzi okuphatha imicimbi: Ngokulayisha izithombe zendawo nokuhlanganisa nezidingo zolimi lwemvelo, ikhiqiza ikhasi lesimemo somcimbi lapho abantu bengabhalisa khona futhi kukhishwe uhlu lwababhalisile.