I-API Ye-MiMo Yombono Engavinjelwe

Iziphetho ezinezindlela eziningi eziyimahhala ziyashibhile ngoba kukhona okunqunyiwe, futhi lokho okunqunyiwe ngokuvamile yingxenye ongeke uyibone: izisindo zicindezelwa zibe ngamabhithi amathathu noma amane, kanti imodeli iyaphendula ngokuzethemba mayelana neshadi engeke isakwazi ukulifunda ngempela. I-MiMo-V2.5 isibonelo esiphambene kule khathalogu yethu — entengweni ephansi kakhulu futhi esigabeni sokunemba esiphakeme kunazo zonke ngesikhathi esisodwa, ngoba isisekelo evela kuso sakhiwe sincane futhi siminyene kunokuba sikhulu futhi singaminyene.

01Lapho i-MiMo ivela khona: imodeli yokucabanga kuqala, imodeli yombono kwesibili

Umzila we-MiMo uwuhlelo locwaningo lwe-Xiaomi olunezisindo ezivulekile, futhi umbono waso wokusungulwa awujwayelekile: ikhono lokucabanga limiswa ngesikhathi sokuqeqeshwa ngaphambili, hhayi lihlonywe kamuva nge-RL. Umsebenzi ikakhulukazi ubunjiniyela bedatha — i-Xiaomi yakha kabusha ukukhishwa nokuhlungwa kombhalo ukuze inyuse ubuningi bamaphethini okucabanga ekwakhiweni kolimi, yakhiqiza idatha yokucabanga eyenziwe ngobuningi obukhulu, futhi yaqeqesha ngaphambili i-MiMo-7B-Base kumathokheni acishe abe izigidigidi ezingu-25 ngaphansi kwenhlanganisela yezigaba ezintathu, ngenjongo eyengeziwe ye-Multiple-Token Prediction esebenza kabili ngesikhathi sokuqagela njenge-decoder ye-speculative (izinga lokwamukelwa liseduze no-90%, ngokusho kuka-Xiaomi).

Ukuqeqeshwa kwangemuva kwenziwe ngamabomu kube lula: izinkinga zezibalo nekhodi ezingu-130,000 ngemivuzo yokunemba esekelwe emithethweni kuphela, ngaphandle kwemodeli yomvuzo efundiwe, ngengqondo yokuthi imodeli yomvuzo ye-neural iyindawo engagengqwa. Imivuzo eqinisekekayo ayikwazi ukuncengwa. I-MiMo-7B-RL ikhipha amanani angahambisani nobukhulu bayo — i-Xiaomi ibika u-68.2% ku-AIME 2024, u-95.8% ku-MATH-500, u-57.8% ku-LiveCodeBench v5 ku-pass@1, ilingana ne-o1-mini ka-OpenAI. Yilesiya isimo semodeli oyibizayo: incane, eqeqeshwe ngokuminyene, ilungiselelwe ukuthi isebenzele inkinga kunokukhumbula impendulo — futhi leso simo sindile kuzizukulwane ezinezindlela eziningi.

02Ukuthi i-MiMo yafunda kanjani ukubona: ukwakhiwa kwe-MiMo-VL

I-MiMo-VL iyigatsha lombono nolimi, futhi ukwakhiwa kwayo kuyafundisa. I-Xiaomi yahlanganisa izingxenye ezintathu: i-encoder ye-ViT yesixazululo semvelo egcina imininingwane ecophelele esikhundleni sokuphoqa isithombe ngasinye kudlule kugridi enesixazululo esiphansi esimisiwe, i-projector ye-MLP eqalwe ngokungahleliwe ibhekise endaweni efihlekile yemodeli yolimi, futhi — okubaluleke kakhulu — uhlaka lokucabanga lwe-MiMo njengemodeli yolimi. Uhlelo lombono lidla ifa le-decoder esivele injengcono ekucabangeni okunezinyathelo eziningi kunalena engcono nje ekubhaleni izihloko.

Ukuqeqeshwa kwaqhutshwa ezigabeni ezine zokuqeqeshwa ngaphambili — ukufudumeza i-projector, ukuhambisana kombono nolimi, ukuqeqeshwa ngaphambili okuvamile okunezindlela eziningi, i-SFT yomongo omude — kumathokheni acishe abe izigidigidi ezingu-2.4, kwase kuba ne-Mixed On-policy Reinforcement Learning (MORL), efaka ukunemba kokuqonda, ukunemba kokusekelwa kombono, ukulunga kokucabanga, kanye nokuthandwa kwabantu/AI emzuliswaneni owodwa we-on-policy. Umbiko uqotho ekutheni ukulungisa konke lokho kanyekanye akuzinzile — ukwethulwa okuqotho kakhulu kunamaphepha amaningi anezindlela eziningi.

Imiphumela eshicilelwe iqinile ngenani lamapharamitha. I-Xiaomi ibika i-MiMo-VL-7B-RL yehlula i-Qwen2.5-VL-7B ku-imisebenzi engu-35 kwengu-40, u-59.4 ku-OlympiadBench (ngenhla kwamamodeli afika ku-78B), no-56.1 ku-OSWorld-G yokusekelwa kwe-GUI, ngaphambili kwama-ejenti esixhumi akhelwe injongo. Ukuvuselelwa kwe-2508 kamuva kwanyusa i-MMMU ngaphezu kuka-70 okokuqala kulolu chungechunge (70.6), yafinyelela ku-94.4 ku-ChartQA, u-92.5 ku-ScreenSpot-v2 no-70.8 ku-VideoMME, futhi yengeza iswishi yokulawula ukucabanga ukuze imikhondo yokucabanga ikwazi ukucishwa ngesicelo ngasinye.

03Kusho ukuthini u-W8A16, nokuthi kungani kubalulekile kakhulu embonweni kunombhalweni

Yonke imodeli esangweni iphethe isijobelelo sokukalwa kwezinga. Akuwona umhlobiso — kukutshela ukuthi yini eyathengiselwana ukuze imodeli ikwazi ukukhonzwa.

Isijobelelo	Yini	Kungani sisetshenziswa
3BIT-REAP	Izisindo zamabhithi amathathu kanye ne-REAP, indlela yokuthena ochwepheshe elahla ochwepheshe abangaqondiswa kaningi kumodeli ye-MoE ngaphambi kokukala izinga	Yona kuphela indlela yokungena nokukhonza izisekelo ezinkulu kunazo zonke ngesivinini esizwakalayo
W4A16-AUTOROUND-REAP	Izisindo zamabhithi amane, ukusebenza kwamabhithi ayi-16, kukalwe izinga nge-AutoRound — ukuyingilisa okufundiwe okubuyisela ukunemba ngaphezu kokuyingilisa okulula okusondele kakhulu — kanye nokuthena kwe-REAP	Indawo ephakathi yezisekelo ezinkulu ezingadingi ukucindezelwa okuqinile kakhulu
W8A16	Izisindo zamabhithi ayi-8, ukusebenza kwamabhithi ayi-16 — isigaba sokwethembeka esiphezulu kunazo zonke esisikhonzayo	Sisetshenziswa lapho isisekelo sincane ngokwanele ukuthi ukucindezelwa okuqinile kungadingekile

I-MiMo-V2.5 izuza umugqa wesithathu ngokuba encane: isisekelo esiminyene kulesi sigaba asidingi ukucindezelwa kwaze kwafika kumabhithi amathathu ukuze singene, ngakho ungakwazi ukugcina ayisi-8, futhi izisindo ozithatha kuzo zihlala ziseduze nalezo eziqeqeshwa. Lokho kubaluleke ngokungalingani emsebenzini wezithombe, ngesizathu sesakhiwo. Ku-decoder yombhalo kuphela, i-embedding exegayo ngaphansi kokukalwa kwezinga ifakwa umongo kabusha yizingqimba zokunaka eziningi zangemuva — amathuba amaningi okubuyisela. Isakhiwo sombono asinaso leso sikhala: ulwazi lombono ludlula ku-encoder eyodwa nakuprojector eyodwa encane ngaphambi kokuba lifike ku-decoder, futhi konke okulandelayo kunqunywa yilokho okukhiqizwe yileso siminyana. Uphawu olucophelele — u-3 uqhathaniswa no-8 ekuskeneni okunokushuqulwa okuphansi, ipikseli eqondile lapho ibha iphela khona ku-axis yeshadi, umkhawulo wesici se-UI ocelwa ukuthi ukhihle kuso — luhlala emininingwaneni yesisindo yobukhulu obuphansi ukukalwa kwezinga okuqinile okuqala ukukulahla.

Imodeli yombono yamabhithi amathathu ayehluleki ngokuzwakalayo; yehluleka ngokuthula, ikhiqize incazelo egeleza kahle yedokhumenti eneminingwane emibili eguqulwe kuyo yonke. Ukunemba kwesisindo okuphezulu kuyinhlangano eshibhile kunazo zonke okuvikela leso kwehluleka, futhi lapha kumahhala — ukhokha intengo ephansi kakhulu noma ngayiphi indlela.

04Intengo ephansi kakhulu, ngokomongo

Amamodeli amane ekhathalogini yethu yamamodeli ayishumi nambili ahlala entengweni ephansi kakhulu engu-$0.50 engenayo / $2.00 ephumayo ngethokheni eyisigidi. Elilodwa kuwo kuphela elithatha izithombe. Nasi lapho amamodeli anombono efika khona kuzindleko zokukhishwayo, inombolo ebusa amabhili amaningi angempela:

Intengo yokukhishwayo ngo-USD ngamathokheni ayisigidi, amanani ohlu lwe-Shannon AI Gateway, Septhemba 2026. Ubude bebha bulingana nemodeli yombono ebiza kunazo zonke ekhathalogini.

Uma kuqhathaniswa ne-Kimi-K3 — imodeli yombono yophahla, futhi eqinile ngempela — i-MiMo-V2.5 icishe ibe shibhile ngokuphindwe ka-9.6 kokukhishwayo futhi shibhile ngokuphindwe ka-7.7 kokungenayo. Lokho akuyona imangalo yokulingana; kungumbono mayelana nesimo somsebenzi. Ezithombeni ezinkulungwane eyikhulu ngenyanga, imodeli yophahla ayikwazi ukutholakala futhi imodeli yezinga eliphansi iyakwazi, futhi umbuzo oqotho ukuthi ingabe imodeli yezinga eliphansi iyeqa umgqa wakho wokunemba — hhayi ukuthi ingabe inqoba isilinganiso.

05I-MiMo-V2.5 kunye ne-MiniMax-M3: amamodeli ombono amabili ashibhile, aqhathaniswa ngokuqotho

Lezi ezimbili zilingana kanye ngentengo, umongo, kanye nobuso besiphetho. Azishintshaniswa: isigaba sokukalwa kwezinga siyisithunywa sokubheja okubili okwehlukene mayelana nokwenza imodeli yombono ibe wusizo.

	MiMo-V2.5 · W8A16	MiniMax-M3 · 3BIT-REAP
Okungenayo / okuphumayo ngesigidi	$0.50 / $2.00	$0.50 / $2.00
Ukunemba kwesisindo	amabhithi ayi-8	amabhithi amathathu + ukuthenwa kochwepheshe
Ubukhulu besisekelo	Incane, iminyene	Enkulu, i-MoE engaminyene
Umbono	Yebo	Yebo
Iwindi lomongo	262,144	262,144
Uhlelo · amathuluzi · ukusakaza	Yebo	Yebo

Khetha i-MiMo-V2.5 lapho isithombe siyingxenye enzima: i-OCR eminyene emiskeneni engaphelele, ukukhishwa kwezinombolo emashadini nasemathebulini, ukusekelwa kwesithombe-skrini ne-UI, ukufunda umbhalo omncane, ukuhlukanisa izithombe zemikhiqizo ezifana kakhulu — noma yini lapho inombolo eguqulwe khona iyisiphosiso hhayi iphutha lokubukeka. Ufuna indlela enokwethembeka okuphezulu kunazo zonke kusuka kuma-pixel kuya kumathokheni, futhi izisindo zamabhithi ayi-8 yindlela ozikuthola ngayo ngale ntengo.

Khetha i-MiniMax-M3 lapho isithombe siyingxenye elula. Isisekelo esikhulu esingaminyene esicindezelwe sibe ngamabhithi amathathu sisagcina ulwazi lomhlaba oluningi kanye nokwanda kokucabanga kunemodeli encane eminyene, ngisho nasemuva kokuthena. Uma umsebenzi u-"bheka le foto bese ucabangela ngokujulile ukuthi isho ukuthini," noma isithombe siyingxenye eyodwa yokungenayo engxoxweni ejwayele ukuba ngumbhalo, isisekelo esikhulu siyithuluzi elingcono — ukubona okucophelele okuwuthengiselene akubona isithiyo esibopha.

06I-MiMo-V2.5 kunye ne-MiMo-V2.5-Pro: umzalwane othengiselene amehlo ngobunzulu

Lokhu ukuqhathaniswa okukhiqiza amathikithi osizo amaningi kunazo zonke, ngakho kufanele sikusho ngokusobala. U-MiMo-V2.5-Pro-W8A16 no-MiMo-V2.5-W8A16 babelana ngentengo, iwindi lomongo, isigaba sokunemba, negama lomndeni. I-Pro ayamukeli izithombe. Akulona uhlu olukhulu olufaka olunye. Kuyisabelo esihlukile sesabelomali esifanayo.

Ukukhonza imodeli enezindlela eziningi kusho ukukhokhela umbhoshongo wombono kanye ne-projector eceleni kwe-decoder — enjongweni yezindleko emisiwe, amandla angasetshenziswanga kumodeli yolimi uqobo. I-Pro isebenzisa konke embhalweni: isisekelo esikhulu esinesikhala esiningi sokubhala okude, ukucabanga okwandisiwe, nekhodi. I-MiMo-V2.5 isebenzisa ingxenye yaso ekubonweni.

Ngakho umthetho wesinqumo umayelana nokungenayo kwakho, hhayi izifiso zakho. Uma noma yisiphi isicelo singafaka isithombe-skrini, ikhasi eliskeniwe, ishadi, idayagremu, noma isithombe, udinga i-MiMo-V2.5 — i-Pro izokwenqaba ibhulokhi yokuqukethwe. Uma sonke isicelo singumbhalo futhi ufinyelele kumodeli yombono ngomkhuba, ushiya ikhwalithi yesisekelo etafuleni ngeze: thatha i-MiMo-V2.5-Pro ngentengo efanayo.

Iphethini efanele ukwebiwa: qondisa ngohlobo lokuqukethwe. Noma yini enengxenye yesithombe iya ku-MiMo-V2.5-W8A16, konke okunye kuya ku-MiMo-V2.5-Pro-W8A16. Umndeni ofanayo, intengo efanayo, isabelomali somongo esifanayo — umqondisi ulula futhi asikho isibalo sezindleko okumele sivumelaniswe.

07Kusho ukuthini ngempela u-"ukuqonda izithombe okungavinjelwe"

Alukho ungqimba lokwenqaba phambi kwale modeli futhi asikho isihlungi sokuqukethwe kokukhishwayo kwayo. Leso yisinqumo somkhiqizo esinomphumela othile emsebenzini wombono: imodeli ichaza okusefreyimini kunokwenqaba ngokusekelwe endabeni.

Lokhu kubalulekile ngoba ama-API ombono ahlungiwe akalungele ngokwesakhiwo uhlu olubanzi lomsebenzi osemthethweni. Ipayipi lokulawula okuqukethwe kumele libheke into elilawulayo bese lisho ukuthi yini; isihlukanisi esingavumi ukuhlola into esihlukanisayo asiyona i-classifier. Lokhu kufana kuhlu olude lwemisebenzi yangempela:

• Amathuluzi okwethenjelwa nokuphepha — ukumaka kanye izithombe i-API evamile engeke izisebenze, okuyizona ezidinga kakhulu ukumakwa.
• Izithombe zezokwelapha, zobufakazi, nezingozi — izithombe zomtholampilo, zokulimala, nezesigcawu izihlungi ezivamile ezizenqabayo njengezinesihloko esibi kungakhathaliseki ukuthi ngubani obuzayo.
• Ucwaningo lokuphepha — ukufunda amakhasi e-phishing, izixhumi ezinobungozi, nezithombe-skrini ezithwetshuliwe ngaphandle kokuthi imodeli izisole ngesisekelo.
• Ukuhlukaniswa kokuqukethwe kwabantu abadala kanye nokufakwa kwemibhalo yomlando — amapayipi akhona ukuthola izinto futhi angeke abheke ngaphandle kwawo, kanye neqoqo lomlando izihlungi ezizibheka njengezingemakwa.

Ukususa ungqimba lokwenqaba kususa isithiyo sobunjiniyela, hhayi esomthetho. I-Shannon AI isetshenziswa yi-Shannon Lab LLC e-New Mexico, USA, futhi ukusetshenziswa kwe-API kubuswa yinqubomgomo yethu yokusebenzisa ngokuzibophezela. Okukhishwayo okungahlungiwe yikhono, hhayi incwadi yemvume.

08Ukuyibiza: iziphetho, ukungenisa izithombe, nokuthi ilingana kuphi

Imodeli u-MiMo-V2.5-W8A16, efinyeleleka ngezilimi ezintathu zezicelo ezikhuluma nezisindo ezifanayo, ngakho ugcina noma iyiphi i-SDK osele unayo:

• /v1/chat/completions — ihambisana ne-OpenAI. Izithombe zingena njengezingxenye zokuqukethwe ze-image_url, noma njenge-URL ye-https noma i-base64 data URI.
• /v1/messages — ihambisana ne-Anthropic. Izithombe zingena njengamabhulokhi okuqukethwe e-image.
• /v1/responses — ulimi lohlobo lwe-Responses.

Ukusakaza nokubiza amathuluzi kuyasebenza kuzo zontathu, ngakho imodeli ingabheka isithombe-skrini bese ibiza umsebenzi mayelana nalokho ekubonile emzuliswaneni owodwa. Izinhlelo ezigcwele zezicelo nezimpendulo zikuse-mibhalweni ye-API, esiphinde sibe nendawo yokudlala ongafaka kuyo isithombe ngaphambi kokubhala noma yikuphi ikhodi.

I-MiMo-V2.5 isekela u-json_schema no-response_format eqinile — umehluko phakathi kwesibonelo nepayipi. Yinike umbhalo wokukhokha kanye nohlelo olunemikhawulo echaziwe yomthengisi, usuku, izinto ngazinye, kanye nesamba, bese uthola into ekhodi yakho engayiqinisekisa kunesigaba okumele usisebenzise nge-regex. Amabili kwamashumi nambili amamodeli kule khathalogu awukho ukusekelwa okugcwele kokukhishwayo okuhlelekile; lena inakho, futhi yingakho isebenza njengesevisi yokukhipha ulwazi ngobuningi obuphezulu.

Iwindi lomongo lingamathokheni angu-262,144; okukhishwayo okukhiqizwayo kukhawulelwe emathokheni angu-16,384 ngesicelo. Izithombe zisebenzisa amathokheni okungenayo njengomsebenzi wesixazululo, ngakho ukuskena kwamakhasi anesixazululo esiphezulu kudla umongo ngokushesha kakhulu kunenani lamagama elilinganayo lombhalo. Hlukanisa amadokhumenti amade ngamakhasi kunokulayisha iskena elinamakhasi angu-300 esicelweni esisodwa, futhi phatha umkhawulo wokukhishwayo njengesithiyo sangempela ekukhishweni okuphelele — hlukanisa amathebula aminyene ngebanga lemigqa kunokuthola impendulo enqunyiwe.

Isilawuli se-reasoning_effort sinamandla angempela lapha ngokunikwa umlando womndeni. Sehlise ekuhlukaniseni ngobuningi obuphezulu, lapho impendulo iyilebula futhi amathokheni okucabanga ayizindleko nje eziphelele; sinyuse ekutolikeni amashadi, izibalo zombono, nokucabanga ngamadayagremu, lapho imodeli kumele ihlanganise izinto eziningana eyizibonayo ngaphambi kokuba iphendule.

Lapho ilingana khona, nalapho ingalingani khona

Okuhambisana kahle. Ukukhishwa kwamadokhumenti nemibhalo yokukhokha ngobuningi obuphezulu; umsebenzi oseduze ne-OCR emiskeneni engaphelele; ukufunda amashadi, amadeshibhodi, namathebula; ukuqonda izithombe-skrini nokusekelwa kwe-UI; ukumaka ikhathalogu ne-QA yombono; amapayipi okulawula okuqukethwe — noma yimuphi umsebenzi onezindlela eziningi lapho izindleko ngekholi ngalinye zibopha futhi ukunemba kwemininingwane ecophelele yiyona into ongafuni ukuyidedela.

Okungahambisani kahle. Ukucabanga okunezindlela eziningi kobunzima bophahla okudinga imodeli enamandla kunazo zonke kungakhathaliseki intengo — umsebenzi we-Kimi-K3, futhi ibiza cishe izikhathi eziyishumi ngesizathu. Ukukhiqizwa kombhalo kuphela, lapho i-MiMo-V2.5-Pro inikeza isisekelo esingcono ngentengo efanayo. Umsebenzi oncike olwazini lomhlaba olubanzi njenge-encyclopedia, lapho isisekelo esikhulu esingaminyene sishaya isisekelo esincane esiminyene. Kanye nanoma yini edinga umongo ongaphezu kuka-262K, okungekho lapha imodeli ekunikezayo.

09Imibuzo ebuzwa njalo

---

Imithombo: Umbiko wobuchwepheshe we-MiMo (arXiv 2505.07608) · Umbiko Wobuchwepheshe we-MiMo-VL (arXiv 2506.03569) · XiaomiMiMo/MiMo-VL · MiMo-VL-7B-RL-2508. Wonke amanani ezilinganiso abikwe yi-Xiaomi wamachekhiphoyinti akhishiwe aqanjwe ngenhla futhi achaza umzila womndeni, hhayi ukwakhiwa oqondile kwalesi siphetho. Amanani entengo, umongo, namafulegi amakhono angamanani ohlu lwe-Shannon AI Gateway, Septhemba 2026.