I-Moonshot AI Ikhulula u-Kimi K2.6 nge-Long-Horizon Coding, I-Agent Swarm Scalling to 300 Sub-Agents kanye nezinyathelo Ezihlanganisiwe ezingu-4,000

I-Moonshot AI, ilebhu ye-AI yase-China ngemuva komsizi we-Kimi, namuhla inemithombo evulekile Kimi K2.6 – imodeli yomdabu ye-multimodal ephusha imingcele yalokho okungenziwa isistimu ye-AI uma iyekwa ukuthi isebenze ngokuzimele ezinkingeni zobunjiniyela besofthiwe eqinile. Ukukhishwa kuqondise ezimweni ezingokoqobo zokusebenzisa: ama-ejenti okubhala amakhodi asebenza isikhathi eside, isizukulwane sokuqala esivela olimini lwemvelo, ama-ejenti afanayo amakhulu ahlanganisa amakhulukhulu ama-sub-ejenti akhethekile ngesikhathi esisodwa, kanye ne-ecosystem entsha evulekile lapho abantu nama-ejenti avela kunoma iyiphi idivayisi behlanganyela emsebenzini ofanayo. Imodeli isiyatholakala manje ku-Kimi.com, ku-Kimi App, i-API, kanye ne-Kimi Code CLI. Izisindo zishicilelwa ku-Hugging Face ngaphansi kweLayisense ye-MIT Elungisiwe.

Hlobo Luni Lwemodeli Lona, Ngobuchwepheshe?

UKimi K2.6 uyi-a Ingxube-yezazi (MoE) imodeli – isakhiwo esivame ukubusa esikalini somngcele. Esikhundleni sokwenza kusebenze yonke imingcele yemodeli kuwo wonke amathokheni ayicubungulayo, imodeli ye-MoE ihambisa ithokheni ngayinye kusethi encane 'yochwepheshe' abakhethekile. Lokhu kukuvumela ukuthi wakhe imodeli enkulu kakhulu ngenkathi ugcina i-inference compute tractable.

I-Kimi K2.6 inezilinganiso eziphelele ze-1 trillion, kodwa kuphela i-32 billion esebenzayo ithokheni ngayinye. Inochwepheshe abangu-384 sebebonke, nabangu-8 abakhethiwe ngethokheni ngayinye, kanye nochwepheshe ohlanganyelwe o-1 ohlala esebenza. Imodeli inezendlalelo ezingu-61 (okuhlanganisa ungqimba olulodwa oluminyene), isebenzisa ubukhulu obufihliwe bokunaka obungu-7,168, ubukhulu obufihliwe be-MoE obungu-2,048 uchwepheshe ngamunye, kanye namakhanda okunaka angu-64.

Ngale kombhalo, i-K2.6 iyi-a i-multimodal yomdabu imodeli – okusho ukuthi umbono ubhakwa ngokwezakhiwo, awubhilidi. Isebenzisa a I-MoonViT isifaki khodi sombono esinamapharamitha angu-400M futhi sisekela okokufaka kwesithombe nevidiyo ngokomdabu. Eminye imininingwane yezakhiwo: isebenzisa I-Multi-head Latent Attention (MLA) njengendlela yokunaka, I-SwiGLU njengomsebenzi wokwenza kusebenze, usayizi wesilulumagama wamathokheni angu-160K, nobude bomongo obungamathokheni angu-256K.

Ngokusatshalaliswa, i-K2.6 iyanconywa ukuthi iqhubeke i-vLLM, I-SGlangnoma I-KTransformers. Yabelana ngezakhiwo ezifanayo njenge-Kimi K2.5, ngakho ukulungiselelwa okukhona kokuthunyelwa kungasetshenziswa kabusha ngokuqondile. Okudingekayo transformers version kuyinto >=4.57.1, <5.0.0.

Izinombolo Zesihlokwana Sokuqoshwa Kwekhodi Okude

I-metric okungenzeka ithole ukunakwa kakhulu emaqenjini e-dev yile I-SWE-Bench Pro — ukuhlola ibhentshimakhi ukuthi imodeli ingakwazi yini ukuxazulula izinkinga ze-GitHub zomhlaba wangempela kumakhosombe wesofthiwe ochwepheshe.

U-Kimi K2.6 uthole amaphuzu angu-58.6 ku-SWE-Bench Pro, uma kuqhathaniswa no-57.7 we-GPT-5.4 (xhigh), 53.4 we-Claude Opus 4.6 (umzamo omkhulu), 54.2 we-Gemini 3.1 Pro (ecabanga phezulu), kanye no-50.7 ku-Kimi K2.5. Ku-SWE-Bench Verified ithola amaphuzu angu-80.2, ihlezi phakathi kwebhande eliqinile lamamodeli ezinga eliphezulu.

Vuliwe I-Terminal-Bench 2.0 kusetshenziswa uhlaka lomenzeli we-Terminus-2, i-K2.6 ifinyelela ku-66.7, uma kuqhathaniswa no-65.4 kukho kokubili i-GPT-5.4 no-Claude Opus 4.6, kanye no-68.5 ku-Gemini 3.1 Pro. Vuliwe I-LiveCodeBench (v6)ithole amaphuzu angu-89.6 vs. Claude Opus 4.6's 88.8.

Mhlawumbe inombolo egqama kakhulu yomthwalo wemisebenzi ye-ejenti yile Ukuhlolwa Kokugcina Kwesintu (i-HLE-Full) ngamathuluzi: I-K2.6 ithola amaphuzu angu-54.0 — ihola yonke imodeli ngokuqhathaniswa, okuhlanganisa i-GPT-5.4 (52.1), Claude Opus 4.6 (53.0), ne-Gemini 3.1 Pro (51.4). I-HLE ibhekwa kabanzi njengelinye lamabhentshimakhi olwazi anzima kakhulu, futhi okuhlukile ngamathuluzi kuhlola ngokukhethekile ukuthi imodeli ingasebenzisa kanjani izinsiza zangaphandle ngokuzenzakalelayo. Ngaphakathi, i-Moonshot ihlola izinzuzo zokubhala amakhodi okude komkhathizwe isebenzisa yazo I-Kimi Code Benchibhentshimakhi yangaphakathi ehlanganisa imisebenzi eyahlukene, eyinkimbinkimbi yokuphela-siya ekupheleni kuzo zonke izilimi nezizinda, lapho i-K2.6 ibonisa ukuthuthuka okuphawulekayo ngaphezu kwe-K2.5.

Ukuthi Kubukeka Kanjani Amahora angu-13 Wokufaka Ikhodi Okuzenzakalelayo Empeleni

Izifundo eziyizici ezimbili zobunjiniyela emqulwini okhishiwe okushiwo 'ukubhala ikhodi ye-long-horizon' ekusebenzeni.

Kowokuqala, u-Kimi K2.6 ulande ngempumelelo futhi wasebenzisa i- Qwen3.5-0.8B imodeli endaweni ku-Mac, yabe isisetshenziswa futhi yathuthukiswa imodeli yokukhomba kuyo Zig – Ulimi oluhle kakhulu lwezinhlelo – olubonisa ukukhiqizwa okujwayelekile okuphuma ngaphandle kokusabalalisa. Kuzo zonke izingcingo zamathuluzi ezingu-4,000+, ngaphezu kwamahora angu-12 okwenziwa okuqhubekayo, nokuphindaphinda okungu-14, i-K2.6 ithuthukise ukusebenza kusuka cishe ku-15 kuya cishe kumathokheni angu-193/isekhondi, ekugcineni yazuza isivinini esicishe sibe ngu-20% ngokushesha kune-LM Studio.

Kweyesibili, uKimi K2.6 walungiswa ngokuzenzakalelayo exchange-coreinjini yokufanisa yezimali eneminyaka engu-8 ubudala. Ngokusebenza kwamahora angu-13, imodeli iphindaphindeke ngamasu okuthuthukisa angu-12, iqala amakholi amathuluzi angaphezu kuka-1,000 ukuze iguqule ngokunembile imigqa yekhodi engaphezu kuka-4,000. Isebenza njengomakhi wezinhlelo onguchwepheshe, i-K2.6 yahlaziya i-CPU futhi yabelwa amagrafu omlilo ukuze ikhombe izingqinamba ezifihliwe futhi yalungisa kabusha i-core thread topology ukusuka ku-4ME+2RE ukuya ku-2ME+1RE — ikhiphe u-185% omaphakathi wokuphuma (kusuka ku-0.43 kuya ku-1.24 MT/3% ukuya ku-3) kanye nokusebenza kwe-from 1. 2.86 MT/s).

I-Agent Swarms: Ukukala Ngokuvundlile, Hhayi Ngokuqondile Nje

Elinye lamakhono e-K2.6 athakazelisa kakhulu ngokwezakhiwo yiwo I-Agent Swarm – indlela yokuqhathanisa imisebenzi eyinkimbinkimbi kuwo wonke ama-ejenti amancane akhethekile, kunokuthembela ochungechungeni lokucabanga olulodwa, olujulile.

I-architecture ikala ngokuvundlile ukuya kuma-ejenti amancane angu-300 asebenzisa izinyathelo ezihlanganisiwe ezingu-4,000 ngesikhathi esisodwa, ukunwetshwa okukhulu kusuka kuma-agent angaphansi kwe-K2.5 angu-100 kanye nezinyathelo ezingu-1,500. Uquqaba lubola ngokuguquguqukayo imisebenzi ibe imisebenzi engaphansi ehlukahlukene – ukuhlanganisa ukusesha okubanzi kwewebhu nocwaningo olujulile, ukuhlaziywa kwemibhalo enkulu ngokubhala kwamafomu amade, kanye nokukhiqizwa kokuqukethwe kwamafomethi amaningi ngokuhambisana – bese iletha imiphumela ehlanganisiwe efaka amadokhumenti, amawebhusayithi, amaslayidi, namaspredishithi ngaphakathi kokusebenza okukodwa okuzimele. Iswebezane lethula ukhonkolo Amakhono amandla: ingaguqula noma iyiphi i-PDF yekhwalithi ephezulu, ispredishithi, isilayidi, noma idokhumenti ye-Word ibe Ikhono elisebenziseka kabusha. I-K2.6 ithwebula futhi igcine i-DNA yedokhumenti yesakhiwo nesitayela, eyivumela ukuthi ikhiqize kabusha ikhwalithi nefomethi efanayo emisebenzini yesikhathi esizayo – cabanga ngakho njengokufundisa iqulu ngesibonelo kunokuba ngokushesha.

Imibukiso ebambekayo ihlanganisa: ukugijima kwe-ejenti engaphansi kwe-100 efane ne-CV eyodwa elayishiwe ngokumelene nezindima ezifanele eziyi-100 e-California futhi ilethe ukuqalisa kabusha okwenziwe ngendlela oyifisayo okungu-100; enye ekhombe izitolo zezitolo ezingama-30 e-Los Angeles ngaphandle kwamawebhusayithi avela ku-Google Amamephu futhi yakha amakhasi okufika ngayinye; kanye nelo eliphendule iphepha le-astrophysics laba ikhono lezemfundo elisebenziseka kabusha lase likhiqiza iphepha locwaningo elinamakhasi angu-40, amagama angu-7,000 eduze kwedathasethi ehlelekile enokufakwa okungu-20,000+ kanye namashadi angu-14 ebanga lesayensi yezinkanyezi.

Use PheqululaComp ibhentshimakhi kumodi ye-Agent Swarm, i-K2.6 ithola u-86.3 uma kuqhathaniswa no-78.4 ku-Kimi K2.5. Vuliwe I-DeepSearchQA (f1-score), i-K2.6 ithola u-92.5 uma iqhathaniswa no-78.6 we-GPT-5.4.

Letha Ama-agent Akho: Amaqembu Ozipho

Ngale kwengqalasizinda ye-Moonshot uqobo, i-K2.6 iyethula Amaqembu Ezipho njengokubuka kuqala kocwaningo – isici esisha esivula i-ejenti yohlobo lwe-architecture ku-ecosystem yangaphandle, ehlukahlukene.

Umgomo oyinhloko wokuklama: ama-ejenti amaningi kanye nabantu basebenza njengabahlanganyeli bangempela endaweni yokusebenza okwabelwana ngayo. Abasebenzisi bangakwazi ukugibela ama-ejenti avela kunoma iyiphi idivayisi, esebenzisa noma iyiphi imodeli, ngamunye ephethe amathuluzi akhe akhethekile, amakhono, nezimo zenkumbulo eziqhubekayo – kungakhathaliseki ukuthi zisetshenziswa kumakhompyutha aphathekayo endawo, amadivayisi eselula, noma izimo zefu. Enkabeni yaleli qulu, i-K2.6 isebenza njengomxhumanisi oguquguqukayo: iqondanisa ngokuguquguqukayo imisebenzi kuma-ejenti ngokusekelwe kumaphrofayili abo wamakhono athile kanye namathuluzi atholakalayo, ithola lapho i-ejenti ihlangabezana nokwehluleka noma izitebele, inikeze kabusha ngokuzenzakalelayo umsebenzi noma ikhiqize kabusha imisebenzi engaphansi, futhi ilawule umjikelezo wempilo ogcwele wokulethwa kusukela ekuqalisweni nasekuqinisekiseni kuze kube sekugcineni.

I-Moonshot ibilokhu isebenzisa i-Claw Groups ngaphakathi ukuze iqhube eyayo ukukhiqiza okuqukethwe kanye nemikhankaso yethula, enabenzeli abakhethekile abahlanganisa Abenzi BeDemo, Abenzi BeBenchmark, Ama-Social Media Agents, kanye Nabenzi Bevidiyo abasebenza ngokufana – ne-K2.6 exhumanisa inqubo. Kuma-devs acabanga ngezakhiwo ze-orchestration enabenzeli abaningi, lokhu kufanelekile ukukubheka: kumelela ukusuka kokuthi 'I-AI ikwenzela wena imisebenzi' kuya kokuthi 'I-AI iqondisa ithimba lama-ejenti ahlukahlukene, amanye owakhile, egameni lakho.'

Ama-ejenti Asebenzayo: Izinsuku Ezi-5 Zokusebenza Okuzenzakalelayo

I-K2.6 ibonisa ukusebenza okuqinile kuma-ejenti aphikelelayo, asebenzayo njenge I-OpenClaw futhi Hermesesebenza kuzo zonke izinhlelo zokusebenza eziningi ngokuqhubekayo, 24/7 ukubulawa. Lokhu kugeleza komsebenzi kudinga i-AI ukuthi ilawule ngokuqhubekayo amashejuli, ikhiphe amakhodi, futhi ihlele ukusebenza kwenkundla ngaphandle kokugadwa umuntu.

Ithimba lengqalasizinda ye-RL ye-Moonshot uqobo lisebenzisa i-ejenti esekelwe i-K2.6 esebenza ngokuzenzakalelayo izinsuku ezi-5, ilawula ukuqapha, ukuphendula kwesigameko, kanye nokusebenza kwesistimu – ebonisa umongo ophikelelayo, ukuphatha umsebenzi onemicu eminingi, kanye nokusebenzisa umjikelezo ogcwele kusukela ekuqwashiseni kuya ekuxazululweni.

Ukusebenza kulo mthetho kukalwa ngokwangaphakathi Ibhentshi likaClawindawo yokuhlola ehlanganisa izizinda ezinhlanu: Imisebenzi Yokubhala Ikhodi, Ukuhlanganiswa Kwe-IM Ecosystem, Ucwaningo Nokuhlaziywa Kolwazi, Ukuphathwa Kwemisebenzi Okuhleliwe, kanye Nokusetshenziswa Kwenkumbulo. Kuzo zonke ezinhlanu, i-K2.6 idlula i-K2.5 ngokuphawulekayo ezilinganisweni zokuqedwa komsebenzi kanye nokunemba kwesicelo samathuluzi – ikakhulukazi ekuhambeni komsebenzi okudinga ukusebenza okuzimele okuqhubekayo ngaphandle kokuqondisa komuntu.

Izindlela Zokusebenza Ezimbili: Ukucabanga Nokushesha

Ngokuhlanganisa ama-devs nge-API, i-K2.6 iyadalula izindlela ezimbili zokucabanga okubalulekile ekuhwebeni kwe-latency/quality:

Imodi yokucabanga yenza kusebenze ukucabanga okugcwele kwechungechunge – imodeli ibangela inkinga ngaphambi kokukhiqiza impendulo yokugcina. Lokhu kunconyelwa ekwenzeni amakhodi okuyinkimbinkimbi nemisebenzi ye-agent, enezinga lokushisa elinconywayo elingu-1.0. Kukhona futhi a londoloza ukucabanga Imodi, egcina okuqukethwe okucabangayo okugcwele kukho konke ukusebenzisana okuhlukahlukene futhi ithuthukisa ukusebenza kuzimo ze-ejenti yekhodi – ikhutshazwe ngokuzenzakalela, kodwa kufanelekile ukuvunyelwa uma ama-ejenti wokwakha adinga ukugcina isimo sokucabanga esihambisanayo ezinyathelweni eziningi.

Imodi esheshayo ikhubaza ukucabanga okunwetshiwe kwezimpendulo zokubambezeleka okuphansi. Ukuze usebenzise i-Instant mode nge-API esemthethweni, dlula {'thinking': {'type': 'disabled'}} phakathi extra_body. Ngokusetshenziswa kwe-vLLM noma kwe-SGlang, dlula {'chat_template_kwargs': {"thinking": False}} esikhundleni salokho, ngezinga lokushisa elinconyiwe elingu-0.6 kanye ne-top-p engu-0.95.

Okuthathwayo Okubalulekile

• I-Kimi K2.6 iyipharamitha engu-1-trillion, imodeli yomdabu ye-MoE ye-multimodal enamapharamitha angu-32B kuphela acushiwe ngethokheni ngayinye, ekhishwe umthombo ovulekile ngokugcwele ngaphansi kweLayisensi ye-MIT Eguquliwe.
• I-K2.6 ihola wonke amamodeli we-frontier ku-HLE-Full ngamathuluzi (54.0), esebenza kahle kakhulu kune-GPT-5.4 (52.1), Claude Opus 4.6 (53.0), kanye ne-Gemini 3.1 Pro (51.4) kwelinye lamabhentshimakhi e-AI anzima kakhulu.
• Ezivivinyweni zomhlaba wangempela, i-K2.6 yashintsha ngokuzenzakalelayo injini yokufanisa yezimali eneminyaka engu-8 ngaphezulu kwamahora angu-13, yalethela u-185% wemali ephakathi nendawo kanye nenzuzo yokusebenza engu-133%.
• I-Agent Swarm ye-architecture ikala kuya kuma-ejenti amancane angu-300 asebenzisa izinyathelo ezihlanganisiwe ezingu-4,000 ngesikhathi esisodwa, futhi ingaguqula noma iyiphi i-PDF, ispredishithi, noma islayidi ibe Ikhono elisebenziseka kabusha eligcina i-DNA yesakhiwo nesitayela.
• I-Claw Groups, eyethulwe njengokuhlola kuqala ucwaningo, ivumela abantu nama-ejenti avela kunoma iyiphi idivayisi esebenzisa noma iyiphi imodeli ukuthi bahlanganyele eqeqebeni elabelwe, ne-K2.6 esebenza njengomxhumanisi oguquguqukayo owabela imisebenzi ngendlela eguquguqukayo, ethola ukwehluleka, futhi elawula imijikelezo yokuphila egcwele yokulethwa.

Hlola I-Model Weights, Ukufinyelela kwe-API futhi Imininingwane yobuchwepheshe. Futhi, zizwe ukhululekile ukusilandela Twitter futhi ungakhohlwa ukujoyina wethu 130k+ ML SubReddit futhi Bhalisela ku Iphephandaba lethu. Linda! ukutelegram? manje ungasijoyina kuthelegramu futhi.

Udinga ukusebenzisana nathi ekuthuthukiseni i-GitHub Repo yakho NOMA Ikhasi Lobuso Lokugona NOMA Ukukhishwa Komkhiqizo NOMA I-Webinar njll.? Xhuma nathi