Qwen 3.8-Max – imodeli yolimi ehamba phambili eyethulwe yiQwen Team yakwa-Aliyun Uyasebenza

I-Qwen 3.8-Max imodeli yolimi ehamba phambili eyethulwe yiQwen Team yakwa-Aliyun, enenani eliphelele lamapharamitha angu-2.4 trillion, eyakhiwe ngokwandisa i-architecture ye-Qwen 3.5. Njengemodeli enamandla kakhulu emndenini wakwa-Qwen kuze kube manje, iyimodeli yokuqala yezinga le-Max okulindeleke ukuthi ivulelwe umphakathi ngama-open-source weights. Imodeli ithuthukiswe ngokuphelele ekubhaleni ikhodi, emsebenzini wasehhovisi, ocwaningweni nasemisebenzini yesikhathi eside. Isekela ukubizwa nge-API yesiteji se-Qianwen AI, ihambisana nama-protocol e-OpenAI ne-Anthropic, futhi ingahlanganiswa ngaphandle kobunzima nama-agent frameworks ajwayelekile afana ne-Claude Code, Codex ne-Qoder.

Ukubhala ikhodi okuzenzakalelayo isikhathi eside: Isekela ukubhala ikhodi okuzenzakalelayo izinsuku eziyishumi, yakha i-harness ezithuthukisa yona futhi iqede ukuqoqwa kwezidingo, ukwabiwa kwama-issue, ukukhiqizwa kwekhodi nokuzilungisa.
Ukuphinda nokudlula izivivinyo zamaphepha ocwaningo: Iphinda izivivinyo zephepha ngokuzimela, bese ihlola imibono engu-18 yokuthuthukisa ngemijikelezo emine yokuzithuthukisa, idlule indlela yokuqala ngamaphuzu angu-2.7 ku-AIME24.
Ukusebenza emincintiswaneni: Yakha futhi ithuthukise isixazululo ngokuzimela kungakapheli amahora angu-24, yenyusa ukunemba kusuka ku-0.60 kuya ku-0.853, futhi yehlule amaqembu abantu abangu-87% abebambe iqhaza.
Ukuthuthukisa ukusebenza kwehhovisi: Ihlanganisa amakhulu ezimo zomsebenzi ezinenani elikhulu, njengommeli ophawula izigaba ezingu-1,284 ngehora elilodwa nomklami okhiqiza i-prototype esebenzisanayo enamakhasi angu-8 ngesikhathi esisodwa.
Ukuthuthukiswa kwamasu e-quantitative: Iqala ngencazelo yomusho owodwa, ihlele cishe ama-sub-agent angu-330, iqede cishe ama-backtest angu-6,000 futhi ilethe isu le-ETF rotation elisuka ekuqaleni liye ekugcineni.
Ukuziklanyela kwama-chip: Ngemijikelezo yokusebenzisana ecishe ibe ngu-500, yehlisa inani lama-gate ku-netlist ye-cryptographic accelerator lisuka ku-8,298 laya ku-678, inciphise indawo ngo-81% futhi ifinyelele ukuvalwa kwesikhathi.

Ukwandiswa kwe-architecture: Yakhiwe phezu kwe-architecture ye-Qwen 3.5, yafinyelela kumapharamitha angu-2.4 trillion esewonke kanye namapharamitha asebenzayo angu-95B, yandisa kakhulu isikali ngenkathi igcina ifa le-architecture.
Uhlelo lwe-reinforcement learning lomhlaba wangempela: Luthuthukisa amakhono omsebenzi ajwayelekile ngokwandisa ngokuhlanganyela inani lezindawo ze-RL namandla okuqeqesha. Lusebenza nezinselele ezintathu ezihlobene: luhlukanisa umsebenzi, indawo yokusebenza kanye ne-Harness ukuze inani lezindawo likhule ngokwemvelo ngendlela ehlanganisayo; lwakha uhlelo lomvuzo oluhlangene oluhlanganisa ukuqinisekiswa okusekelwe ekwenzeni, ukuhlolwa kwe-rubric kombhalo nokukhishwa okubukwayo ngemva kokurenderwa, kanye nokuhlolwa kwe-agentic, ngaleyo ndlela lususe ukungahambisani kwama-verifier akhethelwe umsebenzi ngamunye; lwakha i-online data balancer eyakha kabusha ukusabalalisa okulinganiselayo kwe-batch ngayinye emsebenzini, ebunzimeni, endaweni yokusebenza nase-harness, yehlise i-gradient variance futhi isekele ukwanda okuqhubekayo nokuzinzile kwamandla okuqeqesha.
I-feedback loop yokuzithuthukisa: Imodeli iqhubekisela phambili ukuthuthukiswa ngomshini ozivumelanisa nezimo wokwenza, ukuthola impendulo nokuphindaphinda. Igcina amasu esistimu ahambisanayo kakhulu emijikelezweni yokusebenzisana engamakhulu kuya ezinkulungwaneni, ithembele kuma-feedback loop okulingisa, ukuhlanganisa nokubeka, ukuze yakhe kabusha indlela yedatha ezingeni le-algorithm esikhundleni sokulungisa kancane i-syntax.
Ukuhlelwa okuzenzakalelayo kwe-workflow: Iqhuba ukuhlela umsebenzi ngohlelo, iguqule umqondo wokuhlela ube wuhlelo oluphindaphindekayo, futhi isekele ukushintsha kalula phakathi kocwaningo nentuthuko oluqhubekayo oluxhumanisa uchungechunge olulodwa kanye nokuhlola okukhulu okuhambisana. Lokhu kuguqula ukuqhubekisela phambili imisebenzi eyinkimbinkimbi ngokulandelana kube yi-automation loop engalethwa ngezinga elikhulu ngaphakathi kwengxoxo eyodwa.

Ukubiza ngesiteji se-Qianwen AI: Bhalisa i-akhawunti uthole i-API Key, bese ubiza imodeli ngokuqondile nge-interface ehambisana ne-OpenAI noma i-Anthropic protocol.
Ukulungisa ukujula kokucabanga: Sebenzisa ipharamitha ethi reasoning_effort ukusetha i-xhigh (okuzenzakalelayo, ukuhlaziya okujulile), i-medium (ibhalansi) noma i-low (ukucabanga okusheshayo), ulawule izindleko nokusebenza ngokwezidingo.
Ukuhlanganisa ama-agent framework: Ngemva kokumisa i-API Key ne-Base URL, ingaxhunywa kubasizi bokubhala ikhodi nama-Agent framework ajwayelekile afana ne-Claude Code, Codex, Qoder CLI, Qwen Code ne-OpenClaw.
Ukufaka ama-open-source weights: Kusukela ngesonto elizayo, landa ama-model weights ku-Hugging Face noma ku-ModelScope ukuze uwafake endaweni futhi wenze ukuthuthukiswa okwengeziwe.

Ukusebenza okuzenzakalelayo isikhathi eside kakhulu: Ingagxila ngokuqhubekayo izinsuku ezimbalwa kuya emashumini ezinsuku, izithuthukise nge-feedback loop yokwenza, ukuthola impendulo nokuphindaphinda, ngaphandle kokungenelela komuntu kuyo yonke inqubo.
Ukuqeqeshwa kwe-RL yomhlaba wangempela: Yandisa ngokuhlanganyela inani lezindawo namandla okuqeqesha, futhi ikhombisa ukusebenza okulinganayo nokuqhubeka nokuthuthuka kuma-harness e-QwenWork, Claude Code, Codex, OpenClaw, Hermes nakwamanye.
Amandla okuletha umsebenzi kusukela ekuqaleni kuye ekugcineni: Ngokusebenzisa i-workflow eguqukayo, iguqula umqondo wokuhlela ube wuhlelo oluphindaphindekayo, inciphise imisebenzi eyinkimbinkimbi ebidinga amasonto okuqhubekiselwa phambili ngokulandelana ibe yi-automation loop engalethwa ngezinga elikhulu ngaphakathi kwengxoxo eyodwa.

Iwebhusayithi yephrojekthi: https://qwen.ai/blog?id=qwen3.8

Ukuhlolwa komthetho nokuhambisana nemithetho: Ifunda amakhulu amadokhumenti ngesikhathi esisodwa, imaka izigaba ezifanele ezingaphezu kwenkulungwane kungakapheli ihora elilodwa, ithathe indawo yomsebenzi weqembu lezomthetho ongathatha cishe isonto.
Ukuklanywa kwe-UI/UX prototype: Ikhiqiza ngesikhathi esisodwa i-prototype esebenzisanayo enamakhasi angu-8 enokwethembeka okuphezulu ye-digital banking App, ngaphandle kokulungiswa ngesandla.
Ukuthuthukiswa kwemenyu yokudla: Isebenzisa amakhulu amadokhumenti ezithako ukuze ikhiqize ngesikhathi esisodwa imenyu ephelele enezidlo ezingu-26, igcine izindleko zezithako ku-33.8%.
Ukwakhiwa kwemodeli yobunjiniyela bezakhiwo: Isebenzisa umdwebo owodwa kuphela ukubuyisela imodeli yesakhiwo esivimbela ukuzamazama komhlaba yebhilidi elinezitezi ezingu-30 esipheqululini, futhi isekele ukubuka izinkomba ezibalulekile lapho uhambisa igundane.
Ukuboniswa kwemithi yokuvuselela umzimba: Iguqula ifomu lokuhlola lephepha elingu-2D libe isiboniso esisebenzisanayo se-3D esingazungeziswa ngokukhululeka futhi siveze izingqimba ngayinye ngayinye, kusize iziguli ziqonde indlela yokuvuselela umzimba.