API Yosasidwa ya Masomphenya ya MiMo

Ma endpoint a multimodal otsika mtengo ndi otsika mtengo chifukwa chinachake chinadulidwa, ndipo chimene chimadulidwa nthawi zambiri ndi gawo limene simungachione: zolemera zimapanikizidwa mpaka mabiti atatu kapena anayi, ndipo mtunduwo umayankhabe modzidalira za chart imene sungathenso kuiwerenga kwenikweni. MiMo-V2.5 ndi chitsanzo chotsutsana mu katalogi yathu — pa pansi pa mtengo ndipo pa mlingo wapamwamba wa kulondola nthawi imodzi, chifukwa maziko amene umachokera anamangidwa aang'ono ndi olimba m'malo mwa aakulu ndi osakanikirana.

01Komwe MiMo imachokera: mtundu woganiza choyamba, mtundu wa masomphenya kenako

Mzere wa MiMo ndi pulogalamu ya kafukufuku ya zolemera zotseguka ya Xiaomi, ndipo mfundo yake yoyambitsa siyotchuka: luso la kuganiza limakhazikitsidwa pa nthawi ya kuphunzitsa koyambirira, osati kumangiriridwa pambuyo pake ndi RL. Ntchitoyi makamaka ndi uinjiniya wa deta — Xiaomi inamanganso kutulutsa ndi kusefa mawu kuti ikweze kachulukidwe ka mawonekedwe a kuganiza mu corpus, inapanga deta yopeka ya kuganiza pa mlingo waukulu, ndipo inaphunzitsa MiMo-7B-Base pa pafupifupi matiriliyoni 25 a ma token pansi pa kusakaniza kwa magawo atatu, ndi Multiple-Token Prediction ngati cholinga chothandizira chimene pa nthawi ya kuganiza chimagwiranso ntchito ngati speculative decoder (mlingo wa kuvomereza pafupi ndi 90%, malinga ndi Xiaomi).

Kuphunzitsa kotsatira ndi kosavuta mwadala: mavuto 130,000 a masamu ndi code ndi mphotho za kulondola zozikidwa pa malamulo zokha, palibe mtundu wa mphotho wophunziridwa, potengera kuti mtundu wa mphotho wa neural ndi malo oti azembedwe. Mphotho zotsimikizika sizingakopeke ndi mawu okoma. MiMo-7B-RL imapereka manambala osagwirizana ndi kukula kwake — Xiaomi ikunena za 68.2% pa AIME 2024, 95.8% pa MATH-500, 57.8% pa LiveCodeBench v5 pa pass@1, kufanana ndi o1-mini ya OpenAI. Amenewo ndi mawonekedwe a mtundu umene mukuyitana: wachifupi, wophunzitsidwa molimba, wokonzedwa kudutsa mkati mwa vuto m'malo mokumbukira yankho — ndipo khalidwe limenelo limapulumuka kupita ku mbadwa za multimodal.

02Mmene MiMo inaphunzirira kuona: kamangidwe ka MiMo-VL

MiMo-VL ndi nthambi ya masomphenya ndi chinenero, ndipo kamangidwe kake kamaphunzitsa. Xiaomi inasonkhanitsa zigawo zitatu: ViT encoder ya native-resolution imene imasunga tsatanetsatane wabwino m'malo mokakamiza chithunzi chilichonse kudutsa mu gridi yokhazikika ya resolution yotsika, MLP projector yoyambitsidwa mwangozi kupita mu malo a latent a mtundu wa chinenero, ndipo — chofunika kwambiri — msana wa kuganiza wa MiMo ngati mtundu wa chinenero. Dongosolo la masomphenya limalandira decoder amene ali kale wabwino pa kuganiza kwa masitepe ambiri m'malo mwa amene ali wabwino pa kufotokoza zithunzi zokha.

Kuphunzitsa kunayenda m'magawo anayi a kuphunzitsa koyambirira — kutentha kwa projector, kugwirizanitsa masomphenya ndi chinenero, kuphunzitsa koyambirira kwa multimodal kwathunthu, SFT ya context yaitali — pa pafupifupi matiriliyoni 2.4 a ma token, kenako Mixed On-policy Reinforcement Learning (MORL), kuphatikiza kulondola kwa kuzindikira, kulondola kwa kumangiriza kwa maonekedwe, kulondola kwa kuganiza, ndi zokonda za anthu/AI mu kuzungulira kumodzi kwa on-policy. Lipoti likunena moona kuti kukonza zonsezi nthawi imodzi ndi kosakhazikika — kanenedwe koona kuposa mmene mapepala ambiri a multimodal amaperekera.

Zotsatira zofalitsidwa ndi zamphamvu pa kuchuluka kwa zigawo. Xiaomi ikunena za MiMo-VL-7B-RL kupambana Qwen2.5-VL-7B pa ntchito 35 mwa 40, 59.4 pa OlympiadBench (pamwamba pa mitundu mpaka 78B), ndi 56.1 pa OSWorld-G pa kumangiriza kwa GUI, patsogolo pa ma agent a mawonekedwe omangidwa mwadala. Kukonzanso kwa 2508 kwapambuyo kunakankha MMMU kupitirira 70 koyamba mu mndandandawu (70.6), kunafika pa 94.4 pa ChartQA, 92.5 pa ScreenSpot-v2 ndi 70.8 pa VideoMME, ndipo kunawonjezera switch yolamulira kuganiza kotero kuti mizere ya kuganiza ingazimitsidwe pa pempho lililonse.

03Chomwe W8A16 chikutanthauza, ndi chifukwa chake chikufunika kwambiri pa masomphenya kuposa pa mawu

Mtundu uliwonse pa gateway umanyamula chowonjezera cha quantization. Si chokongoletsa — chimakuuzani chomwe chinasinthanidwa kuti mtunduwo ukhale wotumikirika.

Chowonjezera	Chomwe chili	Chifukwa chake chimagwiritsidwa ntchito
3BIT-REAP	Zolemera za 3-bit kuphatikiza REAP, njira yodula akatswiri imene imachotsa akatswiri omwe amagawiridwa ntchito kawirikawiri mu mtundu wa MoE isanachite quantization	Njira yokhayo yoikira ndi kutumikira maziko akulu kwambiri pa liwiro lomveka
W4A16-AUTOROUND-REAP	Zolemera za 4-bit, zochitika za 16-bit, zopanikizidwa ndi AutoRound — kuzungulira kophunziridwa kumene kumabweza kulondola pamwamba pa kuzungulira ku pafupi mosaganizira — kuphatikiza kudula kwa REAP	Pakati pa maziko akulu amene safunikira kupanikizidwa kwakukulu
W8A16	Zolemera za 8-bit, zochitika za 16-bit — mlingo wa kukhulupirika kwapamwamba kwambiri umene timatumikira	Umagwiritsidwa ntchito pamene maziko ndi ang'onoang'ono mokwanira kuti kupanikiza kwamphamvu kusafunikire

MiMo-V2.5 imapeza mzere wachitatu pokhala yaifupi: maziko olimba a gulu ili safunikira kuphwanyidwa mpaka mabiti atatu kuti akwane, choncho mungathe kusunga asanu ndi atatu, ndipo zolemera zimene mumatengapo zimakhalabe pafupi ndi zimene zinaphunzitsidwa. Izi zimafunika kwambiri pa ntchito ya zithunzi, chifukwa cha kamangidwe. Mu decoder ya mawu okha, embedding imene imasochera pansi pa quantization imapatsidwa context yatsopano ndi zosanjikiza za chisamalira makumi ambiri zotsatira — mwayi wambiri wobwezera. Chulu cha masomphenya chilibe malo otero: chidziwitso cha maonekedwe chimadutsa mu encoder imodzi ndi projector imodzi yopapatiza chisanafike ku decoder, ndipo chilichonse chotsatira chimadalira pa chimene chotchingacho chinapanga. Chizindikiro chaching'ono — 3 motsutsana ndi 8 mu chojambulidwa cha kusiyana kotsika, pikiselo lenileni pamene bar imathera pa mzere wa chart, malire a chinthu cha UI chimene mukupemphedwa kudina — chimakhala mu tsatanetsatane wa zolemera za kukula kotsika umene quantization yamphamvu imachotsa choyamba.

Mtundu wa masomphenya wa 3-bit sulephera mokweza; umalephera mwakachetechete, kupanga kufotokoza koyenda bwino kwa chikalata ndi manambala awiri osinthanitsidwa mu chiwerengero chonse. Kulondola kwapamwamba kwa zolemera ndicho inshuwalansi yotsika mtengo kwambiri yotsutsana ndi kulephera kumeneko, ndipo pano ndi kwaulere — mumalipira mtengo wa pansi mulimonse.

04Mtengo wa pansi, mu context

Mitundu inayi mu katalogi yathu ya mitundu khumi ndi iwiri ili pa pansi pa mtengo pa $0.50 kulowa / $2.00 kutuluka pa miliyoni ya ma token. Umodzi wokha wa iwo umatenga zithunzi. Nawa malo amene mitundu yokhoza masomphenya imagwera pa mtengo wa zotuluka, nambala imene imalamulira ma bilu enieni ambiri:

Mtengo wa zotuluka mu USD pa ma token 1M, mitengo ya mndandanda ya Shannon AI Gateway, Seputembala 2026. Kutalika kwa bar ndi kofanizira ndi mtundu wa masomphenya wokwera mtengo kwambiri mu katalogi.

Motsutsana ndi Kimi-K3 — mtundu wa masomphenya wapamwamba, ndipo wamphamvu zoonadi — MiMo-V2.5 ndi yotsika pafupifupi kuwirikiza 9.6× pa zotuluka ndi kuwirikiza 7.7× pa zolowa. Ichi si chonena cha kufanana; ndi mfundo yokhudza mawonekedwe a ntchito. Pa zithunzi zana masauzande pa mwezi, mtundu wapamwamba sungatheke ndipo mtundu wa pansi ungatheke, ndipo funso loona ndi lakuti mtundu wa pansi umadutsa muyeso wanu wa kulondola — osati ngati umapambana benchmark.

05MiMo-V2.5 motsutsana ndi MiniMax-M3: mitundu iwiri yotsika mtengo ya masomphenya, yoyerekezedwa moona

Awiriwa amafanana ndendende pa mtengo, context, ndi ma endpoint. Sasinthana: mlingo wa quantization ndi chizindikiro cha zoyembekeza ziwiri zosiyana za chimene chimapangitsa mtundu wa masomphenya kukhala wothandiza.

	MiMo-V2.5 · W8A16	MiniMax-M3 · 3BIT-REAP
Kulowa / kutuluka pa 1M	$0.50 / $2.00	$0.50 / $2.00
Kulondola kwa zolemera	8-bit	3-bit + kudula akatswiri
Kukula kwa maziko	Wachifupi, wolimba	Waukulu, MoE yosakanikirana
Masomphenya	Inde	Inde
Zenera la context	262,144	262,144
Schema · zida · streaming	Inde	Inde

Sankhani MiMo-V2.5 pamene chithunzi ndicho gawo lovuta: OCR yolimba pa zojambulidwa zosakwanira, kutulutsa manambala mu ma chart ndi matebulo, kumangiriza kwa zithunzi za skrini ndi UI, kuwerenga mawu ang'onoang'ono, kusiyanitsa zithunzi za zinthu zofanana kwambiri — chilichonse chimene nambala yosinthanitsidwa ndi cholakwika osati kudetsedwa kokongola. Mukufuna njira ya kukhulupirika kwapamwamba kwambiri kuchokera ku ma pikiseli kupita ku ma token, ndipo zolemera za 8-bit ndicho mmene mumazipezera pa mtengo umenewu.

Sankhani MiniMax-M3 pamene chithunzi ndicho gawo losavuta. Maziko akulu osakanikirana opanikizidwa mpaka mabiti atatu akadali ndi chidziwitso cha dziko chochuluka ndi kuganiza kotakata kuposa mmene mtundu wachifupi wolimba ungakhalire, ngakhale pambuyo pa kudula. Ngati ntchito ndi "yang'anani chithunzi ichi ndipo ganizirani mwatsatanetsatane za chomwe chikutanthauza", kapena chithunzicho ndi chilowetso chimodzi mu kukambirana kwa mawu kwambiri, maziko akulu ndi chida chabwino — kuona kolondola kumene munasinthanitsa sikunali chotchinga.

06MiMo-V2.5 motsutsana ndi MiMo-V2.5-Pro: mchimwene amene anasinthanitsa maso ndi kuya

Kuyerekezera kumeneku kumapanga ma tikiti a chithandizo ambiri, choncho ndi kofunika kunena momveka. MiMo-V2.5-Pro-W8A16 ndi MiMo-V2.5-W8A16 zimagawana mtengo, zenera la context, mlingo wa kulondola, ndi dzina la banja. Pro siilandira zithunzi. Si mtundu wophatikiza zonse. Ndi kagawidwe kosiyana ka ndalama zomwezo.

Kutumikira mtundu wa multimodal kukutanthauza kulipira nsanja ya masomphenya ndi projector pambali pa decoder — pa cholinga chokhazikika cha mtengo, mphamvu imene siinaperekedwe ku mtundu wa chinenero wokha. Pro imawononga zonse pa mawu: maziko akulu okhala ndi malo ambiri a kulemba kwa nthawi yaitali, kuganiza kotambasulidwa, ndi code. MiMo-V2.5 imawononga gawo lake pa kuona.

Choncho lamulo la chisankho likukhudza zolowetsa zanu, osati zolinga zanu. Ngati pempho lililonse lingakhale ndi chithunzi cha skrini, tsamba losindikizidwa, chart, chithunzithunzi, kapena chithunzi, mukufuna MiMo-V2.5 — Pro idzakana gawo la zinthulo. Ngati pempho lililonse ndi mawu ndipo munafikira mtundu wa masomphenya mwachizolowezi, mukusiya mtundu wa maziko pa gome popanda chifukwa: tengani MiMo-V2.5-Pro pa mtengo womwewo.

Mawonekedwe oyenera kuba: tumizani malinga ndi mtundu wa zinthu. Chilichonse chokhala ndi gawo la chithunzi chipite ku MiMo-V2.5-W8A16, chilichonse china ku MiMo-V2.5-Pro-W8A16. Banja lomweli, mtengo womwewo, ndalama zomwezo za context — chotumiza ndi chosavuta ndipo palibe chitsanzo cha mtengo choti mugwirizanitse.

07Chomwe "kumvetsa zithunzi kosasidwa" chikutanthauza kwenikweni

Palibe wosanjikiza wa kukana patsogolo pa mtunduwu ndipo palibe chosefera cha zinthu pa zotulutsa zake. Icho ndi chisankho cha chinthu chokhala ndi zotsatira zenizeni pa ntchito ya masomphenya: mtunduwo umafotokoza chimene chili mu chithunzi m'malo mokana potengera mutu.

Izi zimafunika chifukwa ma API a masomphenya osefedwa mwa kamangidwe kake sagwirizana ndi gulu lonse la ntchito zovomerezeka. Njira ya kuyang'anira iyenera kuyang'ana zinthu zimene ikuziyang'anira ndi kunena chimene zili; chowunikira chimene chimakana kuyang'ana chinthu chimene chikuchiika m'gulu si chowunikira. Chomwecho chimagwira pa mndandanda wautali wa ntchito zenizeni:

• Zida za trust and safety — kuika zilembo pa zithunzi ndendende zimene API yodziwika siidzazikonza, zomwe ndi zithunzi zimene zikufunika zilembo kwambiri.
• Zithunzi za zachipatala, za umboni wa milandu, ndi za ngozi — zithunzi za chipatala, za kuvulala, ndi za malo a ngozi zimene zosefera wamba zimakana ngati zowopsa mosasamala kanthu za amene akufunsa.
• Kafukufuku wa chitetezo — kuwerenga masamba a phishing, mawonekedwe oyipa, ndi zithunzi za skrini zolandidwa popanda mtunduwo kukayika za mfundoyo.
• Kuika m'magulu zinthu za akuluakulu ndi kusunga zolemba — njira zimene zilipo kuti zizindikire zinthu ndipo sizingapereke kuona kwa ena, kuphatikiza zosonkhanitsidwa za mbiri yakale zimene zosefera zimaziona ngati zosaikika zilembo.

Kuchotsa wosanjikiza wa kukana kumachotsa chopinga cha uinjiniya, osati cha zamalamulo. Shannon AI imayendetsedwa ndi Shannon Lab LLC ku New Mexico, USA, ndipo kagwiritsidwe ka API kayendetsedwa ndi Ndondomeko yathu ya Kagwiritsidwe Kaudindo. Zotuluka zosasefedwa ndi luso, osati chilolezo.

08Kuyitana: ma endpoint, zolowetsa za zithunzi, ndi pamene ikwanira

Mtunduwo ndi MiMo-V2.5-W8A16, wofikika pa zilankhulo zitatu za mapempho pa zolemera zomwezo, choncho mumasunga SDK iliyonse imene muli nayo kale:

• /v1/chat/completions — yogwirizana ndi OpenAI. Zithunzi zimalowa ngati zigawo za image_url, kaya ngati URL ya https kapena data URI ya base64.
• /v1/messages — yogwirizana ndi Anthropic. Zithunzi zimalowa ngati zigawo za image.
• /v1/responses — chinenero cha mtundu wa Responses.

Streaming ndi kuyitanitsa ntchito zimagwira ntchito pa zonse zitatu, choncho mtunduwo ungathe kuyang'ana chithunzi cha skrini ndi kuyitanitsa ntchito yokhudza chimene wachiona mu mzere umodzi. Ma schema onse a mapempho ndi mayankho ali mu zolemba za API, zimene zilinso ndi malo oyeserera kumene mungamatire chithunzi musanalembe code iliyonse.

MiMo-V2.5 imathandiza json_schema yolimba ndi response_format — kusiyana pakati pa chionetsero ndi njira ya deta. Ipatseni risiti kuphatikiza schema yokhala ndi magawo a mtundu wodziwika a wogulitsa, tsiku, zinthu za mzere, ndi chiwerengero chonse, ndipo mumapeza chinthu chimene code yanu ingachitsimikizire m'malo mwa ndime imene muyenera kuiwerenga ndi regex. Mitundu iwiri mwa khumi ndi iwiri mu katalogi ili ilibe kuthandiza kwathunthu kwa zotuluka zokonzedwa; uwu uli nako, ndipo ndicho chifukwa chake umagwira ntchito ngati utumiki wotulutsa deta wa kuchuluka kwakukulu.

Zenera la context ndi ma token 262,144; zotuluka zopangidwa zili ndi malire a ma token 16,384 pa pempho lililonse. Zithunzi zimadya ma token olowa malinga ndi resolution, choncho zojambulidwa za masamba a resolution yayikulu zimadya context mofulumira kwambiri kuposa mawu a kuchuluka kofanana. Gawani zikalata zazitali m'masamba m'malo moyika chojambulidwa cha masamba 300 mu pempho limodzi, ndipo tengani malire a zotuluka ngati chotchinga chenicheni pa kutulutsa kwathunthu — gawani matebulo olimba malinga ndi mzere m'malo mopeza yankho lodulidwa.

Chosinthira cha reasoning_effort chili ndi mphamvu yeniyeni pano potengera mbiri ya banjali. Chitsitseni pa kuika m'magulu kwa kuchuluka kwakukulu, kumene yankho ndi chizindikiro ndipo ma token a kuganiza ndi mtengo wabodza; chikwezeni pa kutanthauzira ma chart, masamu a maonekedwe, ndi kuganiza kwa zithunzithunzi, kumene mtunduwo uyenera kuphatikiza zinthu zingapo zimene ukuona usanayankhe.

Pamene ikwanira, ndi pamene siikwanira

Kukwanira bwino. Kutulutsa deta mu zikalata ndi marisiti kwa kuchuluka kwakukulu; ntchito yoyandikira OCR pa zojambulidwa zosakwanira; kuwerenga ma chart, ma dashboard, ndi matebulo; kumvetsa zithunzi za skrini ndi kumangiriza kwa UI; kuika zilembo pa katalogi ndi mafunso a maonekedwe; njira za kuyang'anira — ntchito iliyonse ya multimodal kumene mtengo pa yitanidwe umatchinga ndipo kulondola kwa tsatanetsatane ndicho chimene simufuna kusiya.

Kusakwanira. Kuganiza kwa multimodal kwa kuvuta kwapamwamba kofuna mtundu wamphamvu kwambiri mosasamala mtengo — imeneyo ndi ntchito ya Kimi-K3, ndipo imalipiritsa pafupifupi kuwirikiza khumi pa chifukwa. Kupanga kwa mawu okha, kumene MiMo-V2.5-Pro imapereka maziko abwino pa mtengo womwewo. Ntchito yodalira chidziwitso chochuluka cha dziko, kumene maziko akulu osakanikirana amaposa wachifupi wolimba. Ndi chilichonse chofuna context yoposa 262K, imene palibe mtundu pano umapereka.

09Mafunso ofunsidwa kawirikawiri

---

Magwero: Lipoti la ukadaulo la MiMo (arXiv 2505.07608) · MiMo-VL Technical Report (arXiv 2506.03569) · XiaomiMiMo/MiMo-VL · MiMo-VL-7B-RL-2508. Ziwerengero zonse za benchmark ndi zonenedwa ndi Xiaomi za ma checkpoint otulutsidwa otchulidwa pamwambapa ndipo zikufotokoza mzere wa banja, osati kupanga kwenikweni kwa endpoint iyi. Mitengo, context, ndi zizindikiro za luso ndi za mndandanda wa Shannon AI Gateway, Seputembala 2026.