IQuest-Q1 Uyasebenza

IQuest-Q1 iyimodeli eyisisekelo ye-Agent ekhishwe njenge-open source yi-Zhizhi Innovation Research Institute, egxile emisebenzini yekhodi nama-agent. IQuest-Q1 isebenzisa ukwakheka kwe-sparse MoE, inamapharamitha angu-320B isiyonke, kodwa ivula kuphela angu-15B kwithokheni ngayinye, ngaleyo ndlela yehlise izindleko zokucubungula. IQuest-Q1 iqeqeshwe ezindaweni zokwenziwa kusetshenziswa i-harness reinforcement learning eminingi kanye ne-multi-teacher online distillation (MOPD). Ingaxhunywa ku-Claude Code noma ku-Codex, futhi isezingeni eliphezulu emabhodini e-NL2Repo ne-CyberGym, ngokusebenza okulinganiselayo ngokujwayelekile. Iphumelela kakhulu ekukhiqizeni i-frontend, ekuzenzekeleni komsebenzi wasehhovisi nasekuhloleni izinkinga zocwaningo nokuthuthukiswa kwe-AI.

Ukukhiqizwa kwekhodi ezingeni le-repository: Isidingo esichazwe ngomusho owodwa singakhiqiza iphrojekthi ephelele esebenzayo enamafayela amaningi, futhi imodeli isesigabeni esiphambili samamodeli e-open source ebhodini le-NL2Repo.
Ukusebenza kwe-Agent ye-CLI: Iyakwazi ukufunda nokubhala ama-codebase, ukubiza amathuluzi omugqa womyalo, ukuhlola imiphumela nokululama emaphutheni, futhi ixhumane kalula ne-Claude Code ne-Codex.
Ukukhiqizwa kwezinhlelo ze-frontend ezisebenzisanayo: Ingakhiqiza ngesikhathi esisodwa uhlelo lwewebhu olusebenzayo olunezindawo ze-3D, umqondo wokusebenzisana nokwethulwa okubonakalayo, njengemidlalo ye-FPS, imidlalo yezimoto namadolobha angaphansi kolwandle asekelwe kuma-voxel.
Ukuzenzekela okuphelele komsebenzi wasehhovisi: Isebenzisa i-lark-cli ye-Feishu ukuqhathanisa amarekhodi ezingxoxo namadokhumenti, iqedele yonke inqubo yokuhlukanisa izinga lomcimbi, ukubhala umbiko, ukwaba imisebenzi nokulandelela inqubekelaphambili.
Ukuzihlola kocwaningo nokuthuthukiswa kwe-AI: Iyakwazi ukulandelela ukungajwayelekile ekuqeqeshweni kwe-reinforcement learning, ithole imbangela, ilungise ikhodi futhi iqinisekise imiphumela ngamajika amasha ezinkomba.
Ukwenza imisebenzi eyinkimbinkimbi enezinyathelo eziningi: Ibhekana nemisebenzi ephakathi nendawo nenedlula isikhathi eside, iqhubeke ngokuzenzakalela amahora amaningi, ihlanganise izimo ezifana nokuphepha kwenethiwekhi nokusebenza kwamatheminali.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxoxa ngamaphrojekthi e-AI open source
Ukwakheka kwe-sparse MoE: Imodeli isebenzisa i-decoder-only Transformer njengomgogodla kanye nesakhiwo sochwepheshe abaxubile esingavamile. Inamapharamitha angaba ngu-320B isiyonke, kodwa kwithokheni ngayinye ivula cishe amapharamitha angu-15B kuphela.
Inqubo yokuqeqesha ngezigaba: Ukuqeqesha kuhlukaniswe izigaba ezintathu: i-pretraining, ukuqeqesha kwaphakathi nesikhathi kanye ne-post-training. I-pretraining ibeka isisekelo samakhono ajwayelekile nolwazi lomhlaba; ukuqeqesha kwaphakathi nesikhathi kushintsha ukusatshalaliswa kwedatha kugxile kukhodi ne-STEM futhi kufake ama-agent trajectory anomongo omude; i-post-training igxila ku-software engineering, imisebenzi yesikhathi eside nokucabanga okujwayelekile, kusetshenziswa i-SFT ne-reinforcement learning ukuze kuqiniswe imodeli.
Izindawo zokwenziwa kanye ne-multi-Harness reinforcement learning: Ithimba lenza imisebenzi ezindaweni ezisebenzisekayo ezakhiwe ngama-API angempela, amaseva e-MCP nama-repository. Lokhu kuvumela isu elilodwa ukuthi liqeqeshwe kuma-harness amaningi, kugcinwe amathuluzi nezindlela zokuphathwa komongo zayo ngayinye. Ngale ndlela imodeli ifunda ukuxazulula inkinga kanye namaphethini okuziphatha adingekayo ukuze isebenze ezinhlelweni zangempela zama-agent. Ngaphambi kokwenza ukwenziwa ngcono, kwenziwa ukuhlukaniswa kwezimbangela zokwehluleka; amaphutha emodeli uqobo kuphela aba amasiginali okufunda.
I-MOPD multi-teacher online distillation: I-reinforcement learning ikhiqiza amamodeli ochwepheshe ezindaweni ezine: ulwazi lomsebenzisi lwe-agent, ukubambisana kwama-harness amaningi, imisebenzi yesikhathi eside kanye nomsebenzi ojwayelekile wama-agent. I-MOPD isebenzisa imodeli yomfundi eqaliswe nge-SFT checkpoint, efunda kuma-trajectory ayo, ithole amaphuzu ethokheni ngayinye kochwepheshe ohambisanayo, bese ihlanganisa amamodeli ngezigaba ukuze igoqe ochwepheshe abaningi ibe yi-IQuest-Q1 eyodwa.
Inqubo yocwaningo nokuthuthukiswa eziguqukayo ngokubambisana komshini nomuntu: Phakathi kocwaningo nokuthuthukiswa, imodeli ibamba iqhaza ekuzihloleni kwamakhono ayo, ekwakheni izinhlelo zokuqeqesha nasekwenzeni ezinye izivivinyo. Abacwaningi babantu banomthwalo wezinqumo ezibalulekile, njengokulungisa isiqondiso, ukwenza izivivinyo ezibizayo nokwamukela izinguqulo. Ngemuva kokuqinisekiswa komjikelezo ngamunye, izibuyekezo zemodeli, impahla yokuqeqesha nenqubo kugcinwa ukuze kusetshenziswe ngqo emjikelezweni olandelayo, kwakheke iluphu eqhubekayo yokuthuthukisa.

Ukuthola izisindo zemodeli: Landa izisindo zemodeli nekhodi yokucubungula ku-Hugging Face noma ku-GitHub repository; isiteshi se-ModelScope sizotholakala maduze.
Ukusetha isevisi yokucubungula: Sebenzisa i-SGLang noma i-vLLM ukuqalisa isevisi ehambisana ne-OpenAI, edinga i-tensor parallelism yamakhadi angu-8 (TP=8), ukunemba kwe-bfloat16, kanye nokulayishwa kwama-parser akhethekile e-IQuest e-tool-call nase-reasoning.
Ukusheshisa nokulungisa amapharamitha (ongakukhetha): Ukuze uthole i-throughput ephezulu, ungavula i-recursive MTP speculative decoding; ukuze uphinde ukhiqize ukuhlolwa okusemthethweni, kunconywa i-temperature=1.0, i-top-p=0.95 ne-top-k=20.
Ukubiza imodeli ngokuqondile: Thumela izicelo ngokuqondile nge-Python API noma nge-interface ehambisana ne-OpenAI, usebenzise i-chat template yefomethi ye-IQuest ekubizeni amathuluzi nasekucabangeni.
Ukuxhuma kumathuluzi e-Agent: Xhuma ku-Claude Code noma ku-Codex CLI ngesango elisekela ukubizwa kwamathuluzi. Ngemuva kokusetha ikheli lesango ne-API Token, i-IQuest-Q1 ingakwazi ukushayela i-agent ukuze iqedele imisebenzi.

Ukusebenza kokucabanga okusezingeni eliphezulu: Inamapharamitha angu-320B isiyonke, kodwa ivula kuphela angu-15B kwithokheni ngayinye, ithole amandla emisebenzi asezingeni eliphezulu ngezindleko zokucubungula ezisondele kwezomodeli omncane.
Imiphumela elinganiselayo nephelele: Ihlala isezingeni eliphezulu kuwo wonke amabhodi ayisishiyagalombili e-DeepSWE, NL2Repo, CyberGym, Terminal-Bench, JobBench namanye, ngaphandle kobuthakathaka obusobala.
Ikhono lekhodi elivelele: Ekukhiqizeni ezingeni le-repository, ithole amaphuzu angu-63.0 ku-NL2Repo, yeqa i-DeepSeek-V4-Pro ne-Hunyuan Hy4, yaba seqenjini eliphambili lamamodeli e-open source.
Iluphu yangempela ye-Agent: Ayigcini ngokukhiqiza ikhodi kuphela, kodwa iyakwazi ukufunda amalogi, ukusebenzisa amathuluzi nokululama emaphutheni, ilethe imiphumela yokugcina engaqinisekiswa.
Ukuxhuma ngokuchofoza okukodwa kuma-Agent ajwayelekile: Inikeza i-interface ehambisana ne-OpenAI kanye nesithombe esisemthethweni se-Docker, futhi ingashintshwa ngokuqondile ukuze isetshenziswe ku-Claude Code naku-Codex.

Iwebhusayithi esemthethweni yephrojekthi: https://iquestlab.github.io/
GitHub repository: https://github.com/IQuestLab/IQuest-Q1
I-model repository ye-HuggingFace: https://huggingface.co/IQuestLab/IQuest-Q1

Ukuthuthukiswa kohlelo oluhlakaniphile: Ikhiqiza iphrojekthi ephelele esebenzayo enamafayela amaningi ngomusho owodwa, futhi ingafunda ikhodi, ilungise ama-bug futhi iqhube izivivinyo ngokuzenzakalela ku-Claude Code noma ku-Codex, ilethe imiphumela yobunjiniyela engaqinisekiswa.
Ukukhiqizwa kwamakhasi ewebhu nemidlalo esebenzisanayo: Ngokusekelwe encazelweni yombhalo, ikhiqiza ngesikhathi esisodwa izinhlelo ze-frontend ezifana nemidlalo ye-3D, amathuluzi okubona idatha namakhasi okuklama, ihlanganise ukwakhiwa kwesimo, umqondo wokusebenzisana nokwethulwa okubonakalayo.
Ukuzenzekela kwezinqubo zasehhovisi: Isebenzisa i-lark-cli ye-Feishu ukuhlanganisa izingxoxo, amadokhumenti, amashidi nama-meeting, iqedele yonke inqubo yokuqinisekisa ulwazi, yokuhlukanisa izinga lomcimbi, yokubhala umbiko, yokwaba imisebenzi neyokulandelela inqubekelaphambili.
Ucwaningo nokuthuthukiswa kwe-AI kanye nokusebenza kokuqeqesha: Ihlola ngokuzenzakalela ukungajwayelekile ekuqeqeshweni kwemodeli, ithole imbangela futhi ilungise ikhodi, isize ithimba le-AI ekugcineni nasekuthuthukiseni ipayipi lalo locwaningo nokuthuthukiswa.