I-akhawunti esemthethweni ye-“ModelScope Community” yakwa-Aliyun Cloud imemezele ebusuku kakhulu izolo (ngomhla ka-12) ukuthi ithimba le-Alibaba Qwen selivule ngokusemthethweni izisindo zemodeli i-Qwen3.8-2.4T-A95B.
Lesi yisikhathi sokuqala lapho izisindo zemodeli yezinga le-Qwen-Max zivulwa khona: imodeli inamapharamitha angu-2.4T isiyonke, ivula amapharamitha angu-95B ku-Token ngayinye, isekela ngokwendabuko umongo wama-Token angu-262,144, futhi inganwetshwa ifinyelele ku-1,010,000 Token.
I-Qwen3.8 iqhubeka nesakhiwo esixubile se-Qwen3.5, igxile ekuthuthukiseni ikhono lokuqedela imisebenzi ye-Agent yokuhlela, yehhovisi, yocwaningo lwesayensi kanye nemisebenzi yesikhathi eside kusukela ekuqaleni kuze kube sekupheleni. Inguqulo esemthethweni yamafu, i-Qwen3.8-Max, isekelwe kule modeli evulwe izisindo zayo futhi yengeza amakhono engeziwe ezindawo zokukhiqiza.
Ukuhlolwa okusemthethweni kuhlanganisa ama-Agent ohlelo, ama-Agent ajwayelekile, umsebenzi wobungcweti nomongo omude; uma kuqhathaniswa namamodeli afana ne-Opus 4.8, i-Fable 5 ne-GPT 5.6 Sol, imiphumela iyehluka ngokwezindawo, kanti i-PaperBench, i-IFBench namanye ama-Benchmark ithole amaphuzu aphezulu phakathi kwamamodeli abalwe.
Izixhumanisi ezihlobene zimi kanje:
Imodeli: https://modelscope.cn/models/Qwen/Qwen3.8-2.4T-A95B
Ibhulogi: https://qwen.ai/blog?id=qwen3.8
I-Qwen Studio: https://chat.qwen.ai
I-Qwen3.8-2.4T-A95B iyimodeli yolimi eyimbangela, enamapharamitha angu-2.4T isiyonke, futhi evula amapharamitha angu-95B ku-Token ngayinye. Isendlalelo ngasinye se-MoE semodeli siqukethe ochwepheshe abangu-512, i-Token ngayinye ikhetha ochwepheshe abaqondisayo abangu-10, bese ngesikhathi esifanayo ivula uchwepheshe owabiwe ongu-1.
Imodeli iphinde yaqeqeshwa nge-MTP enezinyathelo eziningi (Multi-Token Prediction), okunika izinjini zokucabanga ezisekela le ndlela amandla okubikezela ama-Token amaningi alandelayo.
Ngokwencazelo, ukuqeqeshwa kwangemva kokwakhiwa kwehhovisi ne-Agent kwe-Qwen3.8-Max kufingqwa ngezigaba ezintathu:
Ukunweba ngokuqhubekayo izindawo zangempela nokuzihlukanisa kulezi zilinganiso ezizimele: umsebenzi (Task) (umsebenzi owodwa → imisebenzi eminingi → imisebenzi yezinsuku eziningi), indawo yokusebenza (Workspace) (amafayela amaningi → izinkomba ezinamazinga → izinkomba eziyinkimbinkimbi ezinezinhlobo ezihlukahlukene), kanye ne-Harness (izinhlobo, izinguqulo namakhono ahlukahlukene). Lokhu kwenza inani lezindawo likhule ngokwemvelo ngokuhlanganisa, esikhundleni sokuncika ekuhlanganisweni okwenziwa ngendlela ehlukile ngakunye.
Ukwakha uhlelo olulodwa lwemivuzo oluhlanganisa izindlela zokuqinisekisa ezihlukahlukene zemisebenzi yangempela. Lolu hlelo luhlanganisa ukuhlolwa okusekelwe ekwenzeni, ukuhlolwa kwe-rubric kwemiphumela yombhalo neyokubukwayo eseyenziwe, kanye nokuhlolwa kwe-agentic. Sihlanganisa amafomu nezindlela eziningi zedatha ohlelweni olulodwa lwemivuzo, sinikeze zonke izindawo isignali yomvuzo engaguquki nethembekile futhi sisuse ukungaguquguquki okuvela ngokwemvelo ekugcineni ama-verifier aqondene nomsebenzi ngamunye.
Ukwakha isilinganisi sedatha esiku-inthanethi esakha kabusha i-batch ngayinye ukuze ukusatshalaliswa kwayo kwemisebenzi, ubunzima, izindawo zokusebenza ne-Harness kulingane kakhulu, kwehlise umehluko wama-gradient phakathi kwama-batch, ngaleyo ndlela kusekele ukunwetshwa okuzinzile nokuqhubekayo kwamandla okuqeqesha okufunda ngokuqinisa.