API yeKuona yeMiMo Isina Kudzorwa
Ndiyo modhi yega mukatarogu yedu inosanganisa kururama kwehuremu kwe8-bit nekunzwisisa mifananidzo — uye inodhura chaizvo zvinodhurwa nemamodhi emavara chete ari pasi: $0.50 anopinda / $2.00 anobuda pamatokeni miriyoni.
Pfupiso
MiMo-V2.5-W8A16 ndiyo nhengo inoona yepeya yeMiMo paGedhi reShannon AI, inobva mumutsara weMiMo weXiaomi — modhi yemutauro inotanga nekufungisisa ine shongwe yekuona yakanamirwa pairi — uye inoshandiswa paW8A16, mhando yedu ye quantization ine kutendeseka kwepamusoro kupfuura dzose, pacluster yedu yemaGPU. Context yematokeni 262,144, matokeni 16,384 anobuda akanyanya, kudana zvishandiso, JSON schema yakasimba, kuyerera, nechidzoreso chereasoning_effort. Hapana chidzitiro chekuramba, hapana sefa pane zvinobuda. Inoenzana neMiniMax-M3 pakuona kwakachipa kupfuura kwose mumutsara uye inoikunda pakururama.
Masuwo emidziyo yakawanda akachipa akachipa nokuti pane chakadimburwa, uye chinowanzodimburwa ndicho chikamu chausingaoni: huremu hunodzvanywa kusvika kumabhiti matatu kana mana, uye modhi ichiri kupindura nechivimbo pamusoro pechati chayasingachagoni kuverenga chaizvo. MiMo-V2.5 ndicho chinhu chinopikisa mukatarogu yedu — iri pamutengo wepasi uye pamhando yekururama yepamusoro panguva imwe chete, nokuti hwaro hwainobva hwakavakwa hudiki uye hwakaunganidzana pane hukuru uye hwakapararira.
01Kwainobva MiMo: modhi inofungisisa kutanga, modhi inoona zvichitevera
Mutsara weMiMo ndiwo hurongwa hwetsvakurudzo hwehuremu hwakavhurika hweXiaomi, uye nharo dzawo dzekuvamba hadzifarirwi nevanhu: unyanzvi hwekufungisisa hunotarwa mukati mekudzidziswa kwepakutanga, kwete kunamirwa gare gare neRL. Basa racho rakawanda iunjiniya hwedhata — Xiaomi yakavakazve kubvisa nekusefa mavara kuti ikwidze huwandu hwezvimiro zvekufungisisa mudunhu redhata, yakagadzira dhata rekufungisisa rekunyepedzera pachiyero chikuru, uye yakadzidzisa pakutanga MiMo-7B-Base pamatokeni anenge matiriyoni makumi maviri nemashanu pasi pemusanganiswa wemadanho matatu, ine Multiple-Token Prediction sechinangwa chekubatsira chinoshandawo panguva yekushandisa se speculative decoder (chiyero chekugamuchira chiri pedyo ne90%, maererano naXiaomi).
Kudzidziswa kwemashure kwakaomeswa nemaune: matambudziko emasvomhu neekodhi 130,000 ane mibairo yekururama inotsamira pamitemo chete, pasina modhi yemubairo yakadzidziswa, panharo dzekuti modhi yemubairo yemaneural inzvimbo inogona kubiridzirwa. Mibairo inogona kutariswa haigoni kunyengedzwa. MiMo-7B-RL inoisa nhamba dzisingaenzani nehukuru hwayo — Xiaomi inoshuma 68.2% paAIME 2024, 95.8% paMATH-500, 57.8% paLiveCodeBench v5 papass@1, ichienzana neo1-mini yeOpenAI. Ndicho chimiro chemodhi yaunodana: yakaunganidzana, yakadzidziswa zvakadzama, yakagadzirirwa kushanda nedambudziko pane kuyeuka mhinduro — uye maitiro iwayo anopona achipinda muzvizvarwa zvemidziyo yakawanda.
02Kuti MiMo yakadzidza sei kuona: magadzirirwo eMiMo-VL
MiMo-VL ndiro bazi rekuona-nemutauro, uye kuvakwa kwaro kunodzidzisa. Xiaomi yakaunganidza zvikamu zvitatu: muisi weViT wechiyero chepachawo unochengeta ruzivo rwakatsetseka pane kumanikidza mufananidzo wega wega kupfuura mugirid rechiyero chidiki chakatarwa, muisiri weMLP wakatanga zvisina kurongwa unoenda munzvimbo yakavanzika yemodhi yemutauro, uye — chinokosha — musana wekufungisisa weMiMo semodhi yemutauro. Sisitimu yekuona inogara nhaka yedhikodha yakatonaka pakufunga kwematanho akawanda pane iyo yakangonaka pakunyora tsananguro.
Kudzidziswa kwakaitwa mumatanho mana ekudzidziswa kwepakutanga — kudziya kwemuisiri, kurongeswa kwekuona-nemutauro, kudzidziswa kwepakutanga kwemidziyo yakawanda kwakajairika, neSFT yecontext refu — pamatokeni anenge matiriyoni 2.4, ipapo Mixed On-policy Reinforcement Learning (MORL), ichipeta mukati kururama kwekuona, kururama kwekuratidza nzvimbo, kururama kwekufungisisa, nezvinodiwa nevanhu/AI mudenderedzwa rimwe chete. Gwaro rinotaura pachena kuti kugadzirisa zvose izvozvo panguva imwe chete hakugarisisi — kutaura kwechokwadi kupfuura kunopiwa nemagwaro mazhinji emidziyo yakawanda.
Mhedzisiro yakabudiswa yakasimba pahuwandu hwemapharamita. Xiaomi inoshuma MiMo-VL-7B-RL ichikunda Qwen2.5-VL-7B pamabasa 35 pa40, 59.4 paOlympiadBench (pamusoro pemamodhi anosvika 78B), ne56.1 paOSWorld-G pakuratidza nzvimbo paGUI, iri mberi kwemaagent ezvishandiso zvakavakirwa izvozvo. Chishandurwa che2508 chakazotevera chakasundidzira MMMU kupfuura 70 kekutanga mumutsara uyu (70.6), chikasvika 94.4 paChartQA, 92.5 paScreenSpot-v2 ne70.8 paVideoMME, uye chakawedzera chidzoreso chekufunga kuti tsimba dzekufungisisa dzigone kudzimwa pachikumbiro chimwe nechimwe.
Chinyorwa pamusoro penharo dzevhezheni
Nhamba yega yega iri pamusoro inhamba yakabudiswa naXiaomi pachekipoinzi chakatumidzwa zita, chakabatanidzwa mubhuroko rezvakatorwa riri pazasi. Chinhu chedu chekatarogu ipoinzi yakazotevera mudzinza iroro. Tiri kutsanangura kwainobva modhi nezvakavakirwa mhuri — kwete kutaura kuti zviyero izvozvo zvakananga zvinotamira kuchivakwa chaunodana. Yera basa rako pachako; ndizvo zvinoitirwa mutengo wepasi.
03Kuti W8A16 inorevei, uye sei ichikosha zvakanyanya pakuona kupfuura pamavara
Modhi yega yega paGedhi ine chipfupiso che quantization. Hachisi chishongo — chinokuudza chakachinjaniswa kuti modhi igone kushandiswa.
| Chipfupiso | Chii chacho | Sei chichishandiswa |
|---|---|---|
| 3BIT-REAP | Huremu hwe3-bit pamwe neREAP, nzira yekubvisa nyanzvi inobvisa nyanzvi dzisingawanzotumirwi kubva pamodhi yeMoE quantization isati yaitwa | Ndiyo nzira yega yekuisa nekushandisa hwaro hukuru chaizvo pakumhanya kune musoro |
| W4A16-AUTOROUND-REAP | Huremu hwe4-bit, maactivation e16-bit, akaitwa quantization neAutoRound — kutenderedza kwakadzidzwa kunodzorera kururama pamusoro pekutenderedza kwakapusa kune kuri pedyo — pamwe nekubvisa nyanzvi kweREAP | Nzvimbo yepakati yehwaro hukuru husingadi kudzvanywa kwakanyanya |
| W8A16 | Huremu hwe8-bit, maactivation e16-bit — mhando ine kutendeseka kwepamusoro kupfuura dzose yatinoshandisa | Inoshandiswa apo hwaro hudiki zvekuti kudzvanya kwakaomarara hakudiwi |
MiMo-V2.5 inowana mutsara wechitatu nekuva yakaunganidzana: hwaro hwakazara mukirasi iyi hahudi kudzvanyiridzwa kumabhiti matatu kuti hukwane, saka unogona kuchengeta masere, uye huremu hwaunotora hunosara huri pedyo nehwakadzidziswa. Izvozvo zvinokosha zvakanyanya pabasa remifananidzo, nechikonzero chechimiro. Mudhikodha yemavara chete, embedding inotsauka pasi pe quantization inodzoserwa mucontext nemadanho makumi eattention anotevera — mikana yakawanda yekupona. Chimiro chekuona hachina nzvimbo yakadaro: ruzivo rwezvinoonekwa runopfuura nemuisi mumwe chete nemuisiri mumwe chete wakamanikana rusati rwasvika kudhikodha, uye zvose zviri pasi zvinotarwa nezvakagadzirwa nemukaha iwoyo. Chiratidzo chakatsetseka — 3 kupesana ne8 muskeni isina kupenya, pikiseru chaiyo panopera bhaa paaxis yechati, muganhu wechinhu cheUI chaunobvunzwa kudzvanya — chinogara muruzivo rwehuremu rwehukuru hushoma runotanga kuraswa ne quantization yakaomarara.
Modhi inoona ye3-bit haikundikani nemhirizhonga; inokundikana zvakanyarara, ichigadzira tsananguro yakanaka yegwaro ine manhamba maviri akachinjanisirana muhuwandu hwakazara. Kururama kwehuremu kwakakwira ndiyo inishuwarenzi yakachipa kupfuura dzose pakukundikana ikoko, uye pano ndeyemahara — unobhadhara mutengo wepasi pasinei nezvose.
04Mutengo wepasi, mumamiriro azvo
Mamodhi mana mukatarogu yedu yemamodhi gumi nemaviri ari pamutengo wepasi we$0.50 anopinda / $2.00 anobuda pamatokeni miriyoni. Imwe chete yawo ndiyo inogamuchira mifananidzo. Heino panogara mamodhi anoona pamutengo wezvinobuda, nhamba inotonga mabhiri mazhinji echokwadi:
Mutengo wezvinobuda mumadhora pamatokeni 1M, mitengo yerondedzero yeGedhi reShannon AI, Gunyana 2026. Hurefu hwemutsara hwakaenzaniswa nemodhi inoona inodhura kupfuura dzose mukatarogu.
Kuenzaniswa neKimi-K3 — modhi inoona iri padenga, uye ine simba chairo — MiMo-V2.5 yakachipa ka9.6 pane zvinobuda neka7.7 pane zvinopinda. Idzo hadzisi nharo dzekuenzana; inharo dzichimiro chebasa. Pamifananidzo zviuru zana pamwedzi, modhi yepadenga haitengeki uye modhi yepasi inotengeka, uye mubvunzo wechokwadi ndewekuti modhi yepasi inosvika here pamuganhu wako wekururama — kwete kuti inokunda here muyero.
05MiMo-V2.5 kupesana neMiniMax-M3: mamodhi maviri anoona akachipa, akaenzaniswa nechokwadi
Aya maviri anoenzana chaizvo pamutengo, pacontext, nepachimiro chesuwo. Haasi ekutsivana: mhando ye quantization ichiratidzo chekutenda kuviri kwakasiyana pamusoro pechinoita kuti modhi inoona ibatsire.
| MiMo-V2.5 · W8A16 | MiniMax-M3 · 3BIT-REAP | |
|---|---|---|
| Anopinda / anobuda pa1M | $0.50 / $2.00 | $0.50 / $2.00 |
| Kururama kwehuremu | 8-bit | 3-bit + kubvisa nyanzvi |
| Hukuru hwehwaro | Hwakaunganidzana, hwakazara | MoE huru, yakapararira |
| Kuona | Hongu | Hongu |
| Hwindo recontext | 262,144 | 262,144 |
| Schema · zvishandiso · kuyerera | Hongu | Hongu |
Sarudza MiMo-V2.5 kana mufananidzo uri chikamu chakaoma: OCR yakadzama pamaskeni asina kukwana, kubvisa manhamba pamachati nepatafura, kuratidza nzvimbo pamasikirinishoti nepaUI, kuverenga mavara madiki, kusiyanisa mifananidzo yezvigadzirwa yakafanana chaizvo — chero chinhu apo nhamba yakachinjanisirana chikanganiso pane chinoonekwa chete. Unoda nzira ine kutendeseka kwepamusoro kubva pamapikiseru kuenda kumatokeni, uye huremu hwe8-bit ndiyo nzira yaunohuwana nayo pamutengo uyu.
Sarudza MiniMax-M3 kana mufananidzo uri chikamu chiri nyore. Hwaro hukuru hwakapararira hwakadzvanywa kumabhiti matatu huchiri kubata ruzivo rwenyika rwakawanda nekufara kwekufungisisa kupfuura modhi diki yakazara, kunyange mushure mekubvisa nyanzvi. Kana basa riri rekuti "tarisa mufananidzo uyu ufungisise kwenguva refu pamusoro pezvaunoreva," kana mufananidzo uri chinhu chimwe chinopinda muhurukuro inonyanya kuva yemavara, hwaro hukuru ndicho chishandiso chiri nani — kupinza kwekuona kwawakachinjanisa kwaisava chinhu chinodzivirira.
06MiMo-V2.5 kupesana neMiMo-V2.5-Pro: hanzvadzi yakachinjanisa maziso neudzamu
Uku ndiko kuenzanisa kunogadzira matikiti erubatsiro akawanda kupfuura ose, saka zvakakodzera kutaura zvakajeka. MiMo-V2.5-Pro-W8A16 neMiMo-V2.5-W8A16 zvinogovana mutengo, hwindo recontext, mhando yekururama, nezita remhuri. Pro haigamuchiri mifananidzo. Haisi yakapfuura imwe. Ndiko kugoverwa kwakasiyana kwebhajeti rimwe chetero.
Kushandisa modhi yemidziyo yakawanda kunoreva kubhadhara shongwe yekuona nemuisiri pamwe nedhikodha — pachinangwa chemutengo chakatarwa, unyanzvi husina kushandiswa pamodhi yemutauro pachayo. Pro inoushandisa hwose pamavara: hwaro hukuru hune nzvimbo yakawanda yekunyora kwenguva refu, kufungisisa kwakareba, nekodhi. MiMo-V2.5 inoshandisa chikamu chahwo pakuona.
Saka mutemo wesarudzo uri pazvinopinda zvako, kwete pazvishuwo zvako. Kana chikumbiro chipi zvacho chichigona kuva nesikirinishoti, peji rakaskenwa, chati, dhayagiramu, kana mufananidzo, unoda MiMo-V2.5 — Pro icharamba bhuroko rezvirimo. Kana chikumbiro chega chega chiri chemavara uye wakatora modhi inoona nekujaira, uri kusiya mhando yehwaro pasina chikonzero: tora MiMo-V2.5-Pro pamutengo mumwe chete.
Chimiro chakakodzera kutorwa: tumira zvichienderana nerudzi rwezvirimo. Chero chine chikamu chemufananidzo chinoenda kuMiMo-V2.5-W8A16, zvimwe zvose zvichienda kuMiMo-V2.5-Pro-W8A16. Mhuri imwe cheteyo, mutengo mumwe chete, bhajeti recontext rimwe chete — mutumiri wacho ari nyore uye hapana muenzaniso wemutengo wekuenzanisa.
07Kuti "kunzwisisa mifananidzo kusina kudzorwa" kunorevei chaizvo
Hapana chidzitiro chekuramba mberi kwemodhi iyi uye hapana sefa yezvirimo pane zvinobuda zvayo. Ndiyo sarudzo yechigadzirwa ine chinobva mairi chakatarwa pabasa rekuona: modhi inotsanangura zviri mumufananidzo panzvimbo pekuramba zvichienderana nenyaya yacho.
Izvi zvinokosha nokuti maAPI anoona ane sefa haakodzeri nechimiro chawo kuboka rose rebasa riri pamutemo. Hurongwa hwekudzora zvirimo hunofanira kutarisa zvinhu zvahunodzora hwotaura kuti chii; muongorori anoramba kutarisa chinhu chaanorongedza haasi muongorori. Zvimwe chetezvo zvinoshanda pane rondedzero refu yemabasa echokwadi:
- Zvishandiso zvekuvimbika nekuchengetedza — kunyora marebhuru pamifananidzo chaiyo isingagadziriswi neAPI yakajairika, ndiyo mifananidzo inonyanya kuda marebhuru.
- Mifananidzo yezvokurapa, yeforensic, neyetsaona — mifananidzo yekiriniki, yekukuvara, neyenzvimbo yetsaona inorambwa nemasefa akajairika sekunge inotyisa pasinei nekuti ndiani ari kukumbira.
- Tsvakurudzo yezvekuchengetedza — kuverenga mapeji ephishing, maUI ane hutsinye, nemasikirini akatorwa pasina modhi kunetseka nehwaro hwacho.
- Kurongedza zvirimo zvevakuru nekurongwa kwezvakachengetwa — hurongwa huripo kuti hunzwe zvinhu uye husingakwanisi kupa mumwe basa rekutarisa, pamwe nezvakaunganidzwa zvekare zvinotorwa nemasefa sezvisingakwanisi kunyorerwa marebhuru.
Kubvisa chidzitiro chekuramba kunobvisa chipingamupinyi cheunjiniya, kwete chepamutemo. Shannon AI inoshandiswa neShannon Lab LLC muNew Mexico, USA, uye kushandiswa kweAPI kunotongwa neMutemo wedu Wekushandisa Nemutsigo. Zvinobuda zvisina sefa unyanzvi, kwete tsamba yemvumo.
08Kuidana: masuwo, mifananidzo inopinda, nepainokodzera
Modhi ndiMiMo-V2.5-W8A16, inosvikwa pamitauro mitatu yezvikumbiro inosvika pahuremu humwe chete, saka unochengeta SDK ipi zvayo yaunotova nayo:
/v1/chat/completions— inoenderana neOpenAI. Mifananidzo inopinda sezvikamu zvezvirimo zveimage_url, seURL yehttps kana seURI yedhata yebase64./v1/messages— inoenderana neAnthropic. Mifananidzo inopinda semabhuroko ezvirimo eimage./v1/responses— mutauro werudzi rweResponses.
Kuyerera nekudana zvishandiso zvinoshanda pamitatu yose, saka modhi inogona kutarisa sikirinishoti yodana basa pamusoro pezvayaona mukutenderera kumwe chete. Maschema akazara ezvikumbiro nemhinduro ari mumagwaro eAPI, ayo anochengetawo nzvimbo yekuedza yaunogona kunamira mufananidzo usati wanyora kodhi ipi zvayo.
MiMo-V2.5 inotsigira json_schema yakasimba neresponse_format — musiyano uri pakati pekuratidzira nehurongwa. Ipa risiti pamwe neschema ine minda ine mhando yemutengesi, zuva, zvinhu zviri pamitsetse, nehuwandu hwakazara, uye unowana chinhu chinogona kuongororwa nekodhi yako pane ndima yaunofanira kuparadzanisa neregex. Maviri emamodhi gumi nemaviri ari mukatarogu iyi haana kutsigira kwakazara kwezvinobuda zvakarongwa; iyi ine, uye ndicho chikonzero ichishanda sesevhisi yekubvisa ruzivo yehuwandu hukuru.
Hwindo recontext imatokeni 262,144; zvinobuda zvinogadzirwa zvakaganhurirwa pamatokeni 16,384 pachikumbiro chimwe nechimwe. Mifananidzo inodya matokeni anopinda zvichienderana nechiyero chayo, saka maskeni emapeji echiyero chepamusoro anodya context nekukurumidza kupfuura huwandu hwakaenzana hwemazwi emavara. Kamura magwaro marefu mumapeji pane kuisa sikeni yemapeji mazana matatu muchikumbiro chimwe chete, uye tora denga rezvinobuda semuganhu chairo pakubvisa ruzivo kwakazara — kamura matafura akadzama nemitsara panzvimbo pekuwana mhinduro yakachekwa.
Chidzoreso chereasoning_effort chine simba chairo pano nekuda kwedzinza remhuri. Chidzikise pakurongedza kwehuwandu hukuru, apo mhinduro iri rebheru uye matokeni ekufunga ari mutengo chaiwo; chikwidze pakududzira machati, pamasvomhu ezvinoonekwa, nepakufungisisa pamadhayagiramu, apo modhi inofanira kubatanidza zvinhu zvakati wandei zvainoona isati yapindura.
Painokodzera, nepasingakodzeri
Panokodzera. Kubvisa ruzivo pamagwaro nepamarisiti pahuwandu hukuru; basa rakaita seOCR pamaskeni asina kukwana; kuverenga machati, madheshibhodhi, nematafura; kunzwisisa masikirinishoti nekuratidza nzvimbo paUI; kunyora marebhuru pakatarogu neQA yezvinoonekwa; hurongwa hwekudzora zvirimo — basa ripi zvaro remidziyo yakawanda apo mutengo pakudana kwega kwega uchidzivirira uye kururama kweruzivo rwakatsetseka ndiko kwausingade kurasa.
Pasingakodzeri. Kufungisisa kwemidziyo yakawanda kwakaoma kwepamberi kunoda modhi yakasimba kupfuura dzose pasinei nemutengo — ndiro basa raKimi-K3, uye inodhura kanenge kagumi nechikonzero. Kugadzira mavara chete, apo MiMo-V2.5-Pro inopa hwaro huri nani pamutengo mumwe chete. Basa rinotsamira paruzivo rwenyika rwakawanda, apo hwaro hukuru hwakapararira hunokunda hwakaunganidzana hudiki. Nechero chinhu chinoda kupfuura context ye262K, isingapiwe nemodhi ipi zvayo pano.
Nhamba yega yega yakabudiswa iri pamusoro ndeyaXiaomi pachake, yechekipoinzi chakatumidzwa zita mumutsara weMiMo neweMiMo-VL. Verenga zvinhu zvekutanga pane pfupiso yedu yazvo.
09Mibvunzo inowanzobvunzwa
Chii chinonzi API yekuona yeMiMo isina kudzorwa?
Ndiyo modhi yeMiMo-V2.5 W8A16 paGedhi reShannon AI: modhi inoona iri mudzinza reMiMo reXiaomi, inoshandiswa pacluster yedu yemaGPU pakururama kwehuremu kwe8-bit ine context yematokeni 262,144 uye pasina chidzitiro chekuramba kana sefa pane zvinobuda. Inogamuchira mifananidzo nemavara pa/v1/chat/completions, /v1/messages ne/v1/responses pa$0.50 pamatokeni miriyoni anopinda ne$2.00 pamatokeni miriyoni anobuda.
Ndeipi modhi isina kudzorwa yakachipa kupfuura dzose inonyatsoona mifananidzo?
Mbiri dzinoenzana pamutengo wepasi we$0.50 anopinda / $2.00 anobuda pamatokeni miriyoni: MiMo-V2.5 W8A16 neMiniMax-M3 3BIT-REAP. MiMo-V2.5 ndiyo ine kururama kwepamusoro pane dziri mbiri pahuremu hwe8-bit, sarudzo yakachengeteka pabasa remifananidzo rinoda ruzivo rwakatsetseka. MiniMax-M3 inomhanyisa hwaro hukuru zvikuru hwakadzvanywa kumabhiti matatu, ichifarira kufara kweruzivo rwenyika pane kutendeseka papikiseru imwe neimwe.
Chii chinorehwa neW8A16 uye sei zvichikosha pakuona?
W8A16 zvinoreva huremu hwe8-bit nemaactivation e16-bit, mhando ine kutendeseka kwepamusoro kupfuura dzose mukatarogu yeShannon, inoshandiswa apo hwaro hudiki zvekuti kudzvanya kwakaomarara hakudiwi. Zvinokosha pakuona nokuti shongwe yekuona nemuisiri wayo zvinotakura ruzivo rwakatsetseka mumukaha wakamanikana, uye zvikanganiso zvakaunzwa ipapo hazvidzoswi gare gare senzira inowanzoitika neembedding yemavara isina kujeka zvishoma.
MiMo-V2.5 yakasiyana sei neMiMo-V2.5-Pro?
Mutengo mumwe chete, context imwe cheteyo yematokeni 262,144, mhando imwe cheteyo yeW8A16, asi hadzisi modhi imwe chete. MiMo-V2.5 inogamuchira mifananidzo; MiMo-V2.5-Pro haigamuchiri, ichishandisa bhajeti iroro rekushandisa pahwaro hukuru hwemavara panzvimbo peizvozvo. Shandisa MiMo-V2.5 kana zvinopinda zvichisanganisira masikirinishoti, magwaro, machati kana mifananidzo; shandisa Pro pabasa remavara chete pamutengo mumwe chete.
Modhi inoona yeMiMo isina kudzorwa inotsigira JSON schema nekudana zvishandiso here?
Hongu. MiMo-V2.5 W8A16 inotsigira zvinobuda zvakarongwa zvakasimba zvejson_schema, response_format, kudana zvishandiso, kuyerera nechidzoreso chereasoning_effort. Ndicho chinoita kuti ishandiswe sechigadziriso chekubvisa ruzivo: ipa mufananidzo uwane chinhu chakaongororwa pane mavara aunofanira kuparadzanisa.
Itumire mufananidzo
Kuona kwe8-bit, context ye262K, hapana chidzitiro chekuramba — pamutengo wemodhi yemavara chete.
Verenga Magwaro eAPI Tarisa Mamodhi Ose 12Zita remodhi: MiMo-V2.5-W8A16 · rinotongwa neMutemo wedu Wekushandisa Nemutsigo
Zvakatorwa kubva: gwaro reunyanzvi reMiMo (arXiv 2505.07608) · MiMo-VL Technical Report (arXiv 2506.03569) · XiaomiMiMo/MiMo-VL · MiMo-VL-7B-RL-2508. Nhamba dzose dzemiyero dzakashumwa naXiaomi pamachekipoinzi akaburitswa akatumidzwa mazita pamusoro uye dzinotsanangura dzinza remhuri, kwete chivakwa chaicho chiri pasuwo rino. Mitengo, context, nezviratidzo zveunyanzvi ndizvo zvinhu zverondedzero yeGedhi reShannon AI, Gunyana 2026.