Xiaomi MiMo-V2.6 Uyasebenza

I-Xiaomi MiMo-V2.6 iwuchungechunge lwamamodeli e-multimodal olukhishwe futhi lwaba yi-open source yi-Xiaomi, oluqukethe amamodeli amabili e-multimodal omdabu, i-Pro ne-Flash, olugxile ku-Agentic reinforcement learning yezinga elikhulu egxile emisebenzini eyinkimbinkimbi engaqinisekiswa. Ngemuva kwezinsuku ezi-6 ze-Live RL kanye namatrajectory acishe abe ngu-750,000, le modeli ithuthukisa amakhono obunjiniyela besoftware, okubhala ikhodi, okubona kanye ne-Cyber, futhi inwebela ekwakheni imidlalo ye-3D/Blender, ekulawuleni izingalo zamarobhothi, ku-Computer Use, ekusizeni ucwaningo nasekudaleni amavidiyo nomculo.

Ukucabanga okuyinkimbinkimbi nemisebenzi ye-Agent: Ithuthukiswa nge-reinforcement learning ezimweni eziqinisekisekayo ezihlanganisa ikhodi, imisebenzi evamile, ukubona kanye nokuphepha kwenethiwekhi.
Amakhono obunjiniyela besoftware: Ithuthuke kakhulu ku-DeepSWE v1.1, okuyibhentshimakhi yobunjiniyela besikhathi eside engaphandle kwesampula, kugcizelelwa ukulungisa kwangempela nokwenziwa jikelele.
Ukwakha imihlaba ye-3D: Ihlukanisa izidingo zombhalo/izithombe/amavidiyo zibe yimisebenzi yama-agent amaningi, bese ikhiqiza izigcawu zemidlalo ye-3D ezisebenzayo kanye nomqondo wokusebenzisana.
Ukumodela kwe-3D nge-Blender: Ikhiqiza ama-asset angasetshenziswa ku-animation, ekuphrinteni kwe-3D nasekuthuthukiseni imidlalo ngokusekelwe encazelweni noma esithombeni esiyisithenjwa.
Ukulawula ubuhlakani obugxilile emzimbeni: Isebenzisa izithombe zamakhamera ezibuka ezinhlangothini eziningi njengokufakwayo, bese ilawula ingalo yerobhothi ngeluphu evaliwe ukuze ibambe, ihlele imibala futhi ibeke izinto.
I-Computer Use Agent: Iqonda izixhumi ezibonakalayo futhi isebenzise amathuluzi ehhovisi nawokukhiqiza, iqedele ukusesha, ukuhlela, ukucubungula idatha nokulungisa amaphutha.
Usizo locwaningo: Ingabamba iqhaza “ekuhloleni okomile” kokukhetha izinto zokwakha nasebufakazini bezibalo obusemthethweni ku-Lean 4.
Ukudala okuqukethwe: Ikhiqiza i-frontend, i-PPT, i-SVG, amavidiyo nomculo, kugcizelelwa ubuhle, imiphumela yokunyakaza, ukuhambisana komculo wangemuva nokulandisa.

Landela i-WeChat uphendule ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI open source
I-Agentic RL yezinga elikhulu: Le modeli iqhubeka nokuzama nokulungisa emisebenzini eminingi ye-Code, General, Visual ne-Cyber, isebenzisa imiphumela engaqinisekiswa njengesignali yomvuzo; ngaphansi kwezinsuku ezi-6 iqede amatjectory acishe abe ngu-750,000. Nge-Batch enkulu, i-asynchronous architecture ephelele kanye nokuqeqeshwa okunomongo ongu-1M, isibuyekezo sesinyathelo esisodwa sifinyelela ku-2.7–3.7B token, ngaleyo ndlela kuthuthukiswe ukusebenza kahle kwesampula nokwenziwa jikelele ngaphandle kwesampula.
Indlela yomvuzo nokuvimbela ukukopela: Isebenzisa ukuqhathanisa okuhlobene ngaphakathi kwe-Group ukuze inikeze imisebenzi yesikhathi eside imivuzo enemininingwane, bese isebenzisa ukwakhiwa kwemivuzo, ukuhlolwa okuphikisanayo, ukutholwa kokungajwayelekile kanye nokuqinisekiswa okuhlukahlukene kwama-verifier ukunciphisa i-Reward Hacking; ngesikhathi esifanayo imisa i-MoE Router ukuze ivimbele ukushintsha komthwalo wama-expert futhi iqinisekise ukuzinza kokuqeqeshwa kwezinga elikhulu.
Ukumelwa kwetrajectory okuhlangene nokuhlukaniswa kwezinhlelo: Isebenzisa ukumelwa kwetrajectory okuhlangene kanye nendlela yesijeziso ukwenza amasignali okufunda anemininingwane, isekele ukusebenzisana okuphezulu kohlaka lwama-agent amaningi; ukuhlukanisa i-control plane ne-data plane kwenza kube lula ukudlulisa amatjectory amaningi, kugcine isilinganiso semisebenzi sizinzile kuma-batch axubile, futhi kuthuthukise ukuhambisana kwezinjini zokuqeqesha nezokuqagela.
Ukusuka ku-multimodal kuya ekusebenzeni emhlabeni: Ihlanganisa ulimi, ukubona, ukucabanga ngesikhala se-3D kanye namakhono omdabu okwenza, ukuze le modeli ingagcini ngokukhiqiza ikhodi kuphela, kodwa ikwazi nokulungisa ukuziphatha kwayo ngeluphu evaliwe ngokusekelwe empendulweni ebonakalayo ku-Blender, ezinjini zemidlalo, ezindaweni zokulingisa nakuma-GUI.

Indlela yokuqala: Iklayenti le-MiMo Desktop
Indlela yesibili: I-MiMo Open Platform API
Vula i-XiaoMi MiMo Desktop, uyilande bese ufaka i-MiMo Desktop.
Ngemva kokungena ku-akhawunti, khetha ubulungu bokubhalisa, noma ulungise i-API Key yakho kuzilungiselelo.
Ohlwini lwamamodeli, shintsha uye ku-MiMo-V2.6-Pro / Flash; ezimweni zesikhathi sangempela ungavula i-UltraSpeed.
Sebenzisa ulimi lwemvelo ngokuqondile ukuze uqale imisebenzi yokubhala ikhodi, yokudala i-PPT/ikhasi lewebhu/SVG/ividiyo/umculo, yocwaningo noma yokusebenzisa ikhompyutha, futhi ulayishe izinto ezifana nezithombe, amavidiyo namafayela.
Ngokubuza eminye imibuzo, ngokunikeza impendulo ngezithombe-skrini, ngokucela ukuzama futhi noma ngokuqinisekisa ngezigaba, yenza le modeli ithuthukise futhi ilungise imiphumela ngokusekelwe ekuboneni noma kumalogi.
Ngena ku-Xiaomi MiMo Open Platform, dala uhlelo bese ukopisha i-API Key.
Khetha imodeli i-MiMo-V2.6-Pro / Flash, bese uvula i-UltraSpeed ultra-high-speed mode uma kudingeka.
Ngokwemibhalo ye-Platform API, dlulisa igama lemodeli, imilayezo, amathuluzi/okufakwayo kwe-multimodal kanye nefomethi yempendulo.
Emisebenzini ye-Agent, xhuma amalogi wendawo, izinga lokuphumelela kwezivivinyo, izithombe-skrini noma impendulo ye-verifier ukuze kwakheke iluphu evaliwe “yokwenza—hlola—lungisa”.
Qala ngokuqinisekisa kancane kancane intengo, ukubambezeleka, izinga lempumelelo nokuzinza, bese ushintsha kancane uye emzileni wokukhiqiza.

I-RL ngezinga elikhulu: Igxile emisebenzini eyinkimbinkimbi engaqinisekiswa futhi yenza i-Agentic RL yezinga elikhulu; amatjectory acishe abe ngu-750,000 ezinsukwini ezi-6 anobufakazi bokwenziwa jikelele ngaphandle kwesampula.
Amakhono e-Agent avelele: Iqhathaniswa namamodeli ahamba phambili avaliwe emisebenzini eminingi efana nobunjiniyela besoftware, ama-terminal, amakhasi ewebhu, ukubona kanye ne-Cyber.
Ukusuka ku-multimodal kuya esenzweni: Ihlanganisa umbhalo, izithombe, amavidiyo, ama-GUI, i-3D, i-Blender, izingalo zamarobhothi kanye ne-Computer Use.
Inani elihle kakhulu: I-athikili ithi intengo yezinga elifanayo lobuhlakani icishe ibe ngu-1/20–1/60 yamamodeli aphesheya, futhi intengo ye-API ihlala ingashintshi.
Idizayini ezinzile yobunjiniyela: Isebenzisa ukumiswa kwe-MoE Router, izivikelo ze-Reward Hacking kanye nendlela yetrajectory ehlanganisiwe ukusekela ukuqeqeshwa okuzinzile.
Ilungele kokubili ukudala nocwaningo: Ingakha ama-frontend, ama-PPT, amavidiyo nomculo, futhi ibambe iqhaza ekukhetheni izinto zokwakha nasebufakazini obusemthethweni be-Lean 4.

Ilabhulali yamamodeli ye-HuggingFace：https://huggingface.co/collections/XiaomiMiMo/mimo-v26

I-Agent yobunjiniyela besoftware: Ithola ama-bug ngokuzenzakalelayo, ilungise ikhodi, iqhube izivivinyo, bese iqhubeka nokulungisa ngokusekelwe kumalogi okuhluleka, ilungele ukukhulisa ukusebenza kwe-R&D.
Ukukhiqiza imihlaba esebenzisanayo ye-3D: Ikhiqiza izigcawu zemidlalo emincane ezisebenzayo, ama-asset e-Blender noma ama-prototype Demo kusuka embhalweni/ezithombeni/amavidiyo.
Ukuzenzakalela kwehhovisi nge-Computer Use: Isebenzisa i-GUI ukuqedela ukuseshwa kolwazi, ukuhlelwa kwamaspredishithi, ukuhlelwa kwamadokhumenti kanye nezinqubo eziwela amathuluzi amaningi.
Usizo locwaningo: Yenza ukuhlelwa kwemibhalo nemibhalo yobunikazi, ukukhethwa “kokuhlola okomile” kwezinto ze-MOF kanye nobufakazi obusemthethweni be-Lean 4, phakathi kweminye imisebenzi enzima.
Ukukhiqizwa kokuqukethwe kwe-multimodal: Ikhiqiza ngobuningi amakhasi e-frontend, ama-PPT, ama-SVG, amavidiyo okufundisa kanye nomculo wokulandisa, ukuze kusetshenziswe ekukhangiseni, kwezemfundo nasekukhiqizeni okudala.