NeoHorse-1 Uyasebenza

I-NeoHorse-1 iyimodeli ye-Agent yomthombo ovulekile ekhishwe ngokubambisana yi-Jiyuan Rhythm, i-Wuwen CoreSky, i-Tsinghua, i-Peking University, i-Chinese University of Hong Kong, i-Alibaba nezinye izinhlangano, futhi iza ngezinguqulo ezimbili: i-4B ne-9B. Imodeli isebenzisa ukuqeqeshwa kwangemva kwe-Agent-Native, isebenzisa imizila yokwenziwa yangempela ye-Routing Harness njengedatha yokuqeqesha, iphinde isebenzise izimpawu zomzila njengezisekelo zokuhlela uhlelo lokuqeqesha, ngaleyo ndlela yakhe umjikelezo ovaliwe wempendulo yokuhlola → ukwabiwa kwedatha → ukubuyekezwa kwemodeli. Imodeli isekela ukubiza amathuluzi, ukuhlela imisebenzi, ukusesha okujulile nokukhiqiza ikhodi. Inguqulo ye-9B ithole isilinganiso esingu-69.04, idlula imodeli eyisisekelo ngamaphuzu angu-3.44, isekela ngokwemvelo umongo wamathokheni angu-260,000, futhi ingasatshalaliswa endaweni.

Ukubiza amathuluzi: Ikhetha futhi ibize amathuluzi angaphandle ngokuzenzakalelayo ngokwezidingo zomsebenzi, iqedele imisebenzi enezinyathelo eziningi.
Ukuhlela imisebenzi: Ihlukanisa imisebenzi eyinkimbinkimbi ibe yizinyathelo ezisebenzisekayo, bese yenza ukuhlela nokuphathwa kwesikhathi eside.
Ukusesha okujulile: Icela futhi ihlanganise ulwazi ngenkuthalo phakathi kokuxhumana kwemijikelezo eminingi, isekele uchungechunge lobufakazi lweziphetho.
Ukukhiqiza ikhodi: Ikhiqiza ikhodi esebenzisekayo, iyilungise futhi iyilungise ngokusekelwe empendulweni yendawo kuze kuqedwe umsebenzi.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI omthombo ovulekile
Ukuqeqeshwa kwe-Agent-Native okuqhutshwa imizila ye-Harness: Idatha yokuqeqesha igxile emizileni ehlelekile ekhiqizwa ukwenziwa kwangempela kwe-Routing Harness, ehlanganisa izicelo, imizila, ukubizwa kwamathuluzi, impendulo yendawo nokubuyiselwa emaphutheni. Uma kuqhathaniswa nedatha evamile yemibuzo nezimpendulo, ineza izigaba ezintathu: izidingo zamakhono, izinqumo zokwenziwa kanye nemiphumela yendawo. Imizila ephumelele neyahlulekile iyagcinwa, bese ikhethwa ngemva kokuhlolwa kwesakhiwo nokuhlolwa kwekhwalithi ngezindlela eziyisithupha, ukuze inhloso yokuqeqesha iqondiswe ngqo ekuqedweni komsebenzi.
Ukwabiwa kwedatha okuholwa ngamakhono: Umzila uhlukanisa imisebenzi ibe amazinga amane, C0–C3, ngokwezidingo zamakhono. Izimpawu ziphinde zisetshenziswe ungaxhunyiwe ku-inthanethi njengesisekelo sokuqeqesha: amarekhodi omzila ahlukaniswa abe izingxenye ezintathu, “isibikezelo—isenzo—umphumela”; ingxenye yesibikezelo isetshenziselwa ukuhlela uhlelo lokuqeqesha, kanti ingxenye yomphumela isetshenziselwa ukuthola izikhala zamakhono. Ngale ndlela kulungiswa ukwabiwa kwedatha komjikelezo olandelayo, kwakhiwe umjikelezo ovaliwe wempendulo yokuhlola → ukwabiwa kwedatha → ukubuyekezwa kwemodeli.
Ukuqinisekiswa komjikelezo owodwa kokuzithuthukisa okuphindaphindwayo (RSI): Kuqukethe i-Data-RSI ne-Model-RSI. Eyokuqala isho ukuthi imodeli iqhubeka nokwenza imisebenzi, ikhiqize imizila emisha, bese iqongelela ngokwemvelo izinto zokuqeqesha zakamuva ngemva kokukhetha; eyesibili isho ukubuyekeza imodeli ngokwezikhala ezitholwe ekuhloleni bese iyibuyisela echibini lamamodeli e-Harness. Okwamanje kuqinisekiswe umjikelezo owodwa kuphela; ukwakheka kwezimpawu nemivuzo kusasethelwa abantu.

Faka ithuluzi lokulanda: Sebenzisa i-pip install -U modelscope ukufaka ithuluzi lomugqa womyalo le-ModelScope.
Landa imodeli: Sebenzisa i-modelscope download --model TokenRhythm/NeoHorse-1-9B --local_dir ./NeoHorse-1-9B ukulanda izisindo. Inguqulo ye-4B ilandwa ngendlela efanayo; khetha ngokwesidingo.
Qalisa isevisi: Sebenzisa i-SGLang (pip install "sglang==0.5.17") noma i-vLLM ukuqala isevisi ye-inference, ulungiselele i---context-length 262144, i---reasoning-parser qwen3 kanye ne---tool-call-parser qwen3_coder.
Ukubiza i-API: Sebenzisa i-OpenAI SDK uqondise i-base_url ekhelini lesevisi yendawo, isibonelo http://localhost:8000/v1, bese uthumela izicelo njengokubiza i-OpenAI API.
Qaphela izinsiza: Umongo osebenzisekayo uncike kwimemori ye-GPU. Uma imemori inganele, yehlisa ubude bomongo ngaphambi kokusatshalaliswa.

Ukuqeqeshwa kwe-Agent-Native: Idatha yokuqeqesha ivela ngqo emizileni yokwenziwa yangempela, hhayi kudatha evamile yemibuzo nezimpendulo. Inhloso iqondiswe ekuqedweni komsebenzi, hhayi ekuzivumelaniseni kamuva kwemodeli evamile.
Ukusetshenziswa kwemizila ehlulekile: Isethi yokuqeqesha igcina kokubili amarekhodi aphumelele nawahluleka futhi athathelwa indawo, ukuze imodeli ifunde ukuthi “yikuphi okwenza amaphutha kalula nokuthi ibuyiselwa kanjani ngemva kwephutha”.
Ukuphinde kusetshenziswe izimpawu zomzila: Izimpawu zamazinga ekhono C0–C3 zomzila we-inthanethi ziguqulwa zingaxhunyiwe ku-inthanethi zibe yisisekelo sokuqeqesha, kwakhiwe umjikelezo ovaliwe wempendulo yokuhlola → ukwabiwa kwedatha → ukubuyekezwa kwemodeli.
Usayizi omncane, ukusebenza okuphezulu: Ukusebenza kwe-Agent kwenguqulo ye-4B sekudlule amamodeli amaningi ajwayelekile anemingcele eminingi, kanti inguqulo ye-9B ithole isilinganiso esingu-69.04, idlula imodeli eyisisekelo ngamaphuzu angu-3.44, ngezindleko eziphansi zokusatshalaliswa.
Umongo omude kakhulu: Isekela ngokwemvelo amathokheni angu-262,144, futhi inganwetshwa ibe cishe amathokheni ayizigidi ezingu-1.01, ilungele imisebenzi ye-Agent yesikhathi eside.

Indawo yokugcina ye-GitHub: https://github.com/TokenRhythm/NeoHorse
Ilabhulali yemodeli ye-HuggingFace: https://huggingface.co/collections/TokenRhythm/neohorse-1
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2609.08183

Ukwenziwa kwemisebenzi yama-agent: Isetshenziswa njengemodeli yokwenza ngaphakathi kwama-Harness afana ne-OpenClaw, iqedele ngokuzenzakalelayo imisebenzi yangempela yekhompyutha njengokusebenza ngamafayela nokusebenzisa isofthiwe, iphinde izibuyisele ngokwayo ngemva kokwehluleka.
Ukubiza amathuluzi nokuhlela ama-API: Ikhetha, ihlanganise futhi ibize ama-API angaphandle ngokuzenzakalelayo ezindaweni ezinamathuluzi amaningi, ilungele ukwakha abasizi bezinqubo ezizenzakalelayo.
Ukusesha okujulile nomsizi wocwaningo: Isesha ngenkuthalo emijikelezweni eminingi, iqinisekise ulwazi ngokuphambanisa futhi inikeze uchungechunge lobufakazi, isekele ukukhiqizwa kwemibiko yocwaningo enezinkomba.
Ukuzenzakalela komsebenzi webhizinisi: Iphatha imisebenzi yesikhathi eside enezinyathelo eziningi ezindaweni ezihlukahlukene zomsebenzi ezifana ne-WorkBuddy Bench, njengokulungisa amashejuli, ukuhlela idatha nokukhiqiza imibiko.