Kwekucala, imphendvulo lemfisha: mhla ziyi-22 Inyoni 2026, i-Xiaomi yetfula umndeni wemamodeli i-MiMo-V2.6 futsi yashicilela emasisindvo ngaphansi kwelayisensi ye-MIT, levumela kusetjentiswa kwebhizinisi nekuguculwa cishe ngaphandle kwemibandzela. Lomndeni unemamodeli lamabili: i-Pro, lenemapharamitha langu-1.02 wemathriliyoni, kanye ne-Flash, lenalangu-309 wetigidzigidzi. Kokubili kwemukela umbhalo, titfombe, ividiyo nemsindvo emodelini yinye, ngelifasitelo le-context lelifika emathokhenini lasigidzi sinye. Ku-OpenRouter, i-Flash icala ku-$0.14 ngemathokheni langenako lasigidzi. Ebhizinisini lakho loku kusho tintfo letimbili: lukhetfo lwekuyibamba wena ngekwakho ngelayisensi lelula, kanye nelukhetfo lwe-API lolushibhe kakhulu emisebentini lehlanganisa livi, titfombe nembhalo.

Yini i-Xiaomi leyikhiphe ngempela

Womabili lamamodeli asebentisa luhlaka lwenhlanganisela yetingcweti (MoE), lokusho kutsi yincenye lencane kuphela yemapharamitha lesebentako ethokhenini ngalinye. Yiko loku lokugcina tindleko tekuyisebentisa tingaphansi kakhulu kwaloko bukhulu bayo bonkhe lobungakusikisela:

  • MiMo-V2.6-Pro: emapharamitha langu-1.02 wemathriliyoni sewonkhe, langu-42 wetigidzigidzi lasebentako esinyatselweni ngasinye, kanye ne-context yemathokheni la-1M.
  • MiMo-V2.6-Flash: emapharamitha langu-309 wetigidzigidzi sewonkhe, langu-15 wetigidzigidzi lasebentako, ane-context lefanako.
  • Pro-UltraSpeed: luhlobo lwe-Pro i-Xiaomi letsi lwenta i-inference ngekushesha lokuphindvwe kayi-20, lunikwa nge-API ngentsengo lephindvwe kalishumi.

Inkampani ayigcinanga emasisindvweni. Iphindze yashicilela tindzawo tekucecesha nge-reinforcement learning letingetulu kwa-7,000 letihlanganisa bunjiniyela besofthiwe, kuphindza kwente butsakatsaka bekuphepha, umsebenti welwati kanye nekudizayina ema-web, kanye nesakhiwo sekucecesha ngekwaso, umbiko wetemachwepheshe, nemodeli lencane lekhishwe kulenye lenemapharamitha langu-9 wetigidzigidzi leyakhelwe ku-Qwen. Emasisindvo ashicilelwe nge-FP8 ku-Hugging Face futsi angadawunilodwa ngaphandle kwekucela imvume.

Kusebenta kwayo kumi kuphi ngempela

I-Artificial Analysis inika i-Pro emaphuzu langu-46.32 ku-Intelligence Index yayo. I-Xiaomi itsi lamaphuzu ayibeka embili kwe-Kimi K3 ne-Qwen3.8 Max, loko kuyente kutsi ibe yimodeli yemtfombo lovulekile lenemandla kakhulu ngesikhatsi yetfulwa, nanobe iyavuma kutsi isasemuva kwe-Claude Fable 5.1 ne-GPT-6 Astra.

Kodvwa litafula lekucatsanisa ekhadini lemodeli liyilinganisa nesitukulwane lesendlulile semamodeli lavaliwe, hhayi lesikhona nyalo. Nanti emanani ayo uma icatsaniswa ne-Claude Opus 5:

  • AutomationBench (kwenta imisebenti kutsi iticalele): 53.1 kuya ku-50.3.
  • Terminal Bench 2.1 (umsebenti we-command line): 89.9 kuya ku-89.1.
  • DeepSWE (kulungisa emaphutsa langempela esofthiwe): 71.9 kuya ku-74.0.
  • OSWorld-Verified (kusebentisa ikhompyutha): 82.0 kuya ku-83.4.
  • I-Terminal Bench 4.0 lelukhuni kakhulu: 34.9 kuya ku-49.0.

Sifinyeto lesicotfo: imodeli lesondzele ezingeni lelisetulu lesitukulwane lesendlulile emisebentini yema-agent neyekubhala ikhodi, futsi isemuva ngalokucacile kuleyo lelukhuni kakhulu. Ungasifundzi lesimemetelo njengendlela lelinganako yekungena esikhundleni semamodeli lavaliwe lamasha.

Intsengo: linani leligucula tibalo tekusebenta

Ku-OpenRouter, lisigidzi ngasinye semathokheni langenako nalisigidzi ngasinye semathokheni laphumako kubita:

  • Flash: $0.14 kulangenako na-$0.28 kulaphumako.
  • Pro: $0.435 kulangenako na-$0.87 kulaphumako.
  • Pro-UltraSpeed: $4.35 kulangenako na-$8.70 kulaphumako.

Sibonelo, ngekwemicabango yetfu: umsiti wekusita emakhasimende lophatsa tingcoco leti-1,000 ngelilanga, ngayinye inemathokheni langenako lacishe abe-3,000 kanye nemathokheni laphumako la-500. Loku kuba ngemathokheni langenako latigidzi leti-3 kanye nemathokheni laphumako laluhhafu wesigidzi ngelilanga. Nge-Flash, tindleko ticishe tibe ngu-$0.56 ngelilanga, cishe ngu-SAR 63 ngenyanga. Nge-Pro, cishe ngu-$1.74 ngelilanga, cishe ngu-SAR 196 ngenyanga. Emanani langempela atawuntjintja kuya ngebude betingcoco takho kanye ne-context caching, kodvwa lokubalulekile bukhulu balinani: tindleko akusiso sitsiyo sekuyilinga. I-Xiaomi itsi i-Pro ibita emkhatsini wencenye yemashumi lamabili nencenye yemashumi lasitfupha yemamodeli langaphandle lasezingeni lelifanako lekuhlakanipha; leso simangalo senkampani, hhayi kulinganisa lokutimele.

Imodeli yinye levako, lebonako nalefundzako

Incenye lefanele kunakwa kutsi imodeli yinye iyaticondza tonkhe tinhlobo letine tekufakwako. Lamuhla, nawufuna luhlelo loluhlatiya kuchwetjwa kwetingcingo temakhasimende, ujwayele kudzinga kucala insita yekuguculela inkhulumo ibe ngumbhalo, bese kuba yimodeli yelulwimi lehlatiya lombhalo. Kunjalo nangesitfombe se-invoyisi lesitsetfwe ngumtsengisi ngelucingo: insita yinye yekukhipha umbhalo esitfombeni, bese kuba yimodeli yekuwucondza. Lichungechunge ngalinye kuloluchungechunge lisho tindleko letengetiwe kanye nalenye indzawo lapho kungahluleka khona.

Imodeli lesebentisa tindlela letinyenti ifinyeta lolu chungechunge: kuchwetjwa kwelucingo kungena njengoba kunjalo, sitfombe serisidi njengoba sinjalo, sicephu sevidiyo lesivela kukhamera yesitolo njengoba sinjalo. Loku kwenta kube melula kwakha tinhlelo letinjengekubuyeketa likhwalithi yetingcingo, kufaka ngekutentakalela imibhalo ohlelweni lwakho lwe-accounting, noma kuhlola titfombe tetimphahla nativa. Khumbula, nobe kunjalo, kutsi imitfombo ayisho lutfo lolucondzile ngekusebenta kwayo nge-Arabhu noma ngetilimi temmango taseSaudi, futsi yiko kanye loko lokufanele ukuhlole ngemininingwane yakho ngaphambi kwekutsatsa nobe ngusiphi sincumo.

Kuyisebentisa kumaseva akho: kungenteka, kodvwa udzinga ibhajethi

Ilayisensi ye-MIT isho kutsi ungayisebentisa lemodeli ngekhatsi kwengcalasizinda yakho, uyigucule futsi uyifake emkhicitweni wekuhweba. Loku kubalulekile kunobe ngubani lophatsa imininingwane yemuntfu, ngobe Umtsetfo Wekuvikela Imininingwane Yemuntfu waseSaudi Arabia ubeka imitsetfo ekudluliselweni kwemininingwane ngaphandle kweMbuso, kantsi kuyibamba wena ngekwakho kuyigcina ngekhatsi endzaweni yakho.

Kodvwa bukhulu babo bungempela. Likhadi lemodeli ye-Pro likhombisa kuhleleka kwesibonelo lokuhlukanisa imodeli ku-GPU letingu-8 nge-vLLM, noma kuma-node lamabili nge-SGLang. I-Flash idzinga kuhlukaniswa ku-GPU leti-4. Likhadi alisho luhlobo lwe-GPU, kodvwa imodeli lenkhulu kangaka ngempela idzinga emaseva e-AI emazinga lasetulu. Kungako i-Flash, nemapharamitha ayo langu-15 wetigidzigidzi lasebentako, ilikhandidethi leliliciniso lekuyibamba wena ngekwakho emabhizinisini lamanyenti. Sichaze tibalo te-hardware ngalokuningiliziwe ku-indlela yekusebentisa imodeli lenkhulu ye-AI emshinini wakho.

Yini lokufanele uyinake ngaphambi kwekutsi uyetsembe

  • Kuphepha: i-Pro isemuva kakhulu ekuhlolweni kwekusebentisa butsakatsaka, itfola emaphuzu langu-47.9 ku-ExploitBench uma icatsaniswa nala-70.0 e-Opus 5. Nangabe umsebenti wakho kubuyeketa kuphepha kwekhodi, akusilo lukhetfo lwakho lwekucala.
  • Indlela yemininingwane: kuyisebentisa nge-API kusho kutsi imininingwane yemakhasimende akho idlula kumaseva emhlinzeki wangaphandle. Kumininingwane lebucayi, kuyibamba wena ngekwakho noma kufihla emagama ngaphambi kwekutfumela kuyimfuneko, hhayi lukhetfo.
  • Emanani emtsengisi: imiphumela leminyenti yekuhlolwa yashicilelwa yi-Xiaomi ngekwayo. Yihlole emsebentini lovela emsebentini wakho, ngemibuto lesuselwa emininingwaneni yakho.

Indlela lesiyisingatsa ngayo e-Origami

Asitibophi tinhlelo lesitakhako kumodeli yinye. Sifaka lungcweti lwasemkhatsini kute imodeli ikhone kushintjwa ngaphandle kwekwakha kabusha i-application, bese silinganisa likhandidethi ngalinye emsebentini wangempela lovela emsebentini welikhasimende: kunemba, tindleko ngakunye kusebenta kanye nesikhatsi sekuphendvula. I-MiMo-V2.6 seyijoyina lolu luhlu lolufishane nyalo, ikakhulukati emisebentini lehlanganisa livi, titfombe nembhalo, nakunobe nguyiphi indzawo lapho sikweletu senyanga kuyiko lokuvimba kutsi liphrojekthi le-AI licale. Nangabe unemsebenti lonjalo futsi ufuna kwati tindleko tawo tangempela ngaphambi kwekutsi utibophelele, sitsintse.

Imitfombo