I-API ye-MiniMax engacinezelwanga: umbono ngexabiso elisezantsi

Phantse yonke i-API yeendlela ezininzi ebanjelweyo ibeka umbono njengento ephezulu. Uhlawula umlinganiselo wemodeli yombhalo ngombhalo, uze uhlawule ngaphezulu — ngamanye amaxesha izihlandlo ezininzi ngaphezulu — kwangoko umfanekiso ungena kumyalelo. Ngaphezu koko, imifanekiso engenayo kulapho ukuziphatha kokwala kunamandla khona: imodeli eza kuxoxa ngovuyo ngomxholo ngombhalo iya kwala ukuchaza into enye efotweni. Umphumo yimveliso ebiza kakhulu kumbizo ngamnye kwaye engathembekanga kumbizo ngamnye, nto leyo ingoyena mahluko mbi xa ucubungula umthamo. I-MiniMax-M3 kwi-Shannon yimpendulo yethu kuloo nto: i-$0.50 / $2.00 efanayo oyihlawulela imodeli yethu yombhalo etshiphu kunazo zonke, isetyenziswa kwisiseko esaqeqeshwa ngeendlela ezininzi ukususela kwinyathelo lokuqala, ngaphandle komaleko wokwala ohleli phambi kwaso.

01Ithetha ntoni ngokwenene "eyona modeli yombono engacinezelwanga itshiphu" apha

Ikhathalogu ye-Shannon ziimodeli ezilishumi elinambini. Ezintlanu kuzo zamkela imifanekiso. Amaxabiso azo, kwitoken enye yezigidi:

Imodeli enombono	Engenayo	Ephumayo	Imveliso ecwangcisiweyo
MiniMax-M3 · 3BIT-REAP	$0.50	$2.00	json_schema + response_format
MiMo-V2.5 · W8A16	$0.50	$2.00	json_schema + response_format
Kimi-K2.6 · W4A16-AUTOROUND-REAP	$0.78	$3.67	json_schema + response_format
inkling · W4A16-AUTOROUND-REAP	$1.42	$6.07	response_format kuphela
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	json_schema + response_format

I-MiniMax-M3 ihlala ezantsi — $0.50 / $2.00 — ilingana ne-MiMo-V2.5 yaye akukho nto ingaphantsi kwayo, kuba i-$0.50 / $2.00 ngumgangatho osezantsi woluhlu lwethu lwamaxabiso. Xa ithelekiswa nomphezulu woluhlu lombono, itshiphu izihlandlo eziyi-7.7 kwezingenayo nezihlandlo eziyi-9.6 kwezimphumayo kune-Kimi-K3.

Yibeke leyo emsebenzini. Masithi ucubungula imifanekiso eliwaka — imifanekiso yesikrini, amaphepha askeniweyo, iifoto zeemveliso — ephakathi kwiitoken ezimalunga ne-1,200 zomfanekiso nomyalelo ezingenayo neetoken ezingama-300 zempendulo eziphumayo:

Amaxabiso oluhlu lwe-Shannon, Septemba 2026. Amanani eetoken ngawomzekelo; ukwahlulwa kweetoken zomfanekiso wakho kuxhomekeke kwisisombululo nakwinani lemifanekiso oyincamathelisayo kumbizo ngamnye.

Amanqaku amabini ereferensi angaphezulu awafanele ukugcinwa engqondweni. Ukuzalisa wonke umxholo weetoken ezingama-262,144 kanye kubiza malunga neesenti ezili-13. Impendulo yobude obuphezulu yeetoken ezingama-16,384 ibiza malunga neesenti ezi-3.3. Yiyo leyo imida yokwenene yesicelo esinye esibi kunazo zonke, kwaye mincinci ngokwaneleyo ukuze isithintelo esinomdla kumgca wokusebenza wombono siyeke ukuba yimali size sibe ngumsinga.

02Umnombo we-MiniMax: ilebhu elungiselela iitoken, hayi izihloko

I-MiniMax AI yilebhu yaseShanghai apho yonke imbali yayo yobunzima obuvulekileyo imalunga nombuzo omnye: uyenza njani ingqwalasela ibe tshiphu ngokwaneleyo ukuze umxholo omde neendlela ezininzi ziyeke ukuba zizinto zobutyebi? Ukukhutshwa ngakunye kolu tsala yimpendulo kuloo mbuzo, kwaye iimpendulo zakhelana.

MiniMax-01 — i-Lightning Attention, nombono ukususela ekuqaleni

Uthotho lwe-MiniMax-01 (olwapapashwa njenge-arXiv 2501.08313) lwazisa i-Lightning Attention, ukusetyenziswa okuqaphela i-I/O kohlobo lwengqwalasela ethe tye, ngaphakathi kwimodeli edibeneyo ye-Mixture-of-Experts yeeparamitha ezingama-456B zizonke ezinezingama-45.9B ezisebenzayo kwitoken nganye. Uthotho lwahamba njengesibini: i-MiniMax-Text-01 kunye ne-MiniMax-VL-01, imodeli yombono. I-VL-01 yasebenzisa isitaki se-ViT-MLP-LLM esaziwayo ngoku, kodwa ne-Vision Transformer yeeparamitha ezizizigidi ezingama-303 eyaqeqeshwa kwasekuqaleni kwizibini zomfanekiso-nesihloko ezizizigidi ezingama-694 endaweni ye-encoder ebolekiweyo, yaza yalungelelaniswa yaza yalungiswa kunye kumgca wokusebenza wamanqanaba amaninzi ocubungula iitoken ezizigidigidi ezingama-512.

Loo nkcukacha ibalulekile kwinqaku olifundayo. Umbono kulo mnombo ayisosandiso esiza emva esixotyelelwe kwimodeli yombhalo ukuze kufakwe uphawu. I-MiniMax ibiqeqesha inqaba yayo yokubona ukususela ekukhutsheni kwayo kokuqala okuvulekileyo, yiyo loo nto ukuqonda amaxwebhu, ukuqonda i-UI nemifanekiso yesikrini, nokufunda iitshati kuye kwaba ngamandla ahleli enjalo osapho endaweni yokuba zizinto ezicingwe emva kwexesha.

MiniMax-M1 — ingqwalasela edibeneyo yokuqiqa ixesha elide

I-MiniMax-M1 (arXiv 2506.13585) yachazwa ngababhali bayo njengeyona modeli yokuqiqa yokuqala evulekileyo, enkulu, enengqwalasela edibeneyo. Yagcina umzimba we-MoE we-456B / 45.9B yaza yaphithikeza iindidi ezimbini zengqwalasela: ibhloko ye-transformer yengqwalasela ye-softmax eqhelekileyo emva kwebhloko nganye zesixhenxe ze-Lightning Attention. Umvuzo walinganiswa ngezibalo endaweni yendawo kuluhlu — i-MiniMax yaxela ukuba kubude bokuvelisa beetoken ezingama-100K, i-M1 yasebenzisa malunga nama-25% ee-FLOPs ze-DeepSeek R1, ngoxa ixhasa ngokwemvelo umxholo weetoken ezisigidi.

MiniMax-M2 — ukujika kokusebenza okuncinci

I-MiniMax-M2 yahamba ngakwelinye icala ngobungakanani yaza yazuza ingqwalasela ngenxa yoko: i-decoder yeemaleko ezingama-62 eneeparamitha ezingama-229.9B zizonke kodwa ezingama-9.8B kuphela ezisebenzayo kwitoken nganye, isebenzisa iingcali ezicokisekileyo ezingama-256 kwimaleko nganye ye-MoE ngokurawuta kwe-sigmoid ukuya kwezisibhozo kuzo, iindawo ze-RoPE ne-QK-RMSNorm ngozinzo loqeqesho. Yayiyimodeli yokukhowuda ye-agentic ngokucacileyo — izilungiso zeefayile ezininzi, imijikelo yokuqhuba-nokulungisa, izilungiso eziqinisekiswe ngovavanyo — kwaye yazisa ifomathi yokucinga okuphithikezelweyo, apho ukuqiqa kuthwalwa phakathi kwemijikelo yokubiza izixhobo endaweni yokulahlwa emva komjikelo ngamnye. I-M2 yimodeli eyamisela udumo lwe-MiniMax lokuhambisa ukuziphatha kwe-agentic okukufuphi kumphambili kwisiqingatha sexabiso letoken nganye, ngokugcina ukusebenza kuncinci kuphela.

MiniMax-M3 — isiseko esiphantsi kwale endpoint

I-MiniMax-M3 yakhutshwa nge-1 Juni 2026, ubunzima bufike kwi-Hugging Face kungekudala emva koko kunye nengxelo yobuchwepheshe kwi-arXiv 2606.13392. Yimodeli ye-Mixture-of-Experts yeeparamitha ezingama-428B enezingama-23B ezimalunga ezisebenzayo kwitoken nganye, inomxholo weetoken ezisigidi obhaliweyo, kunye nemifanekiso neevidiyo ezingenayo zemvelo — i-MiniMax iyichaza njengeqeqeshwe ngeendlela ezixubeneyo "ukususela kwinyathelo lokuqala", endaweni yokuqeqeshwa kwangaphambili ngombhalo ize ilungiselelwe iindlela ezininzi emva koko. Ngamanani ayo apapashiweyo i-MiniMax ifumana ama-80.5% kwi-SWE-bench Verified kunye nama-59.0% kwi-SWE-bench Pro. Lawo ngamanani axelwe ngumthengisi kubalo-mgangatho lokukhowuda, nto leyo ikanye indlela ofanele ukuwathelekelela ngayo; siyawacaphula kuba ngamabango apapashiweyo elebhu, kungekhona kuba siwavelise kwakhona ngokuzimeleyo.

03I-MiniMax Sparse Attention: kutheni imodeli yombono inokuba tshiphu kangaka

Ubunjineli obunomdla kwi-M3 yi-MiniMax Sparse Attention (MSA), kwaye sisizathu esithe ngqo sokuba imodeli yeendlela ezininzi yobu bungakanani inokuhlala kwixabiso lethu elisezantsi. Imifanekiso ibiza iitoken ezininzi. Amaphepha ambalwa esisombululo esiphezulu anokufaka amashumi amawaka eetoken kumyalelo phambi kokuba ubhale umyalelo omnye. Phantsi kwengqwalasela eqhelekileyo, elo xabiso landa ngokuphindwe kabini, kwaye imodeli efunda amaphepha angama-20 ayibizi ngokuphindwe kabini kunaleyo ifunda alishumi — ibiza kakhulu ngaphezu kokuphindwe kabini.

I-MSA itshintsha umilo waloo jikeleze. Yindlela esasazekileyo ngeebhloko eyakhiwe kwisiseko se-Grouped Query Attention, yahlulwe yaziimpazi ezimbini:

• Isebe le-Index Branch elilula elinika amanqaku iibhloko ze-key-value — amanqaku afunyenwe ngokudibanisa oyena mkhulu kwiitoken ezingaphakathi kwebhloko nganye — lize likhethe ezimbalasela ezingu-k kwindawo nganye yombuzo nakwiqela ngalinye le-GQA, lisoloko libandakanya ibhloko yalapha equlethe umbuzo ngokwawo ngozinzo loqeqesho.
• Isebe le-Main Branch eliqhuba ingqwalasela eqhelekileyo ye-softmax kuphela kwiibhloko ezikhethwe yindeksi.

Ucwangciso kwiphepha le-MSA yi-k = iibhloko ezili-16 zeetoken ezili-128, into enika umbuzo ngamnye uhlahlo-lwabiwo-mali olumisiweyo lweetoken ezingama-2,048 ezinikwe ingqwalasela. Olumisiweyo, oko kuthetha alukhuli nobude bolandelelwano. Lelo ngamaqhinga onke: njengoko umxholo ukhula, ixabiso lengqwalasela yomlinganiselo oshinyeneyo lihlala linyuka ngoxa ixabiso lesebe eliphambili le-MSA lihlala lilingana, ngoko ingqwalasela ikhula ngokuthe tye endaweni yokukhula ngokuphindwe kabini.

Amanani avela kwingxelo ye-MiniMax Sparse Attention (arXiv 2606.13392): ukwehla okuphindwe izihlandlo eziyi-28.4 kwii-FLOPs zengqwalasela kwitoken nganye xa kuthelekiswa ne-GQA ecwangciswe ngokufanayo kumxholo weetoken ezisigidi, kunye nokukhawuleza kwexesha lewotshi okuphindwe izihlandlo eziyi-14.2 kokuzalisa / neziyi-7.6 kokukhupha okulinganiselwe kwizixhobo ze-H800 kuphando lweeparamitha ezingama-109B kwelo phepha. I-MiniMax ixela ngokwahlukileyo ukukhawuleza okuphindwe izihlandlo eziyi-9 kokuzalisa neziyi-15 kokukhupha kwi-M3 xa kuthelekiswa ne-M2 kumxholo weetoken ezisigidi. Onke axelwe yilebhu.

Iphepha likwenza nokujonga okubalulekileyo kwinqaku lombono: kwi-MoE ye-109B eneendlela ezininzi zemvelo, i-MSA yagcina ukusebenza okulinganayo nomlinganiselo wengqwalasela epheleleyo kwiibalo-mgangatho zemifanekiso kubandakanya i-AI2D, i-ChartQA ne-MMMU, nakwiibalo-mgangatho zeevidiyo kubandakanya i-EgoSchema, i-LongVideoBench ne-VideoMME — apho iindidi ezisasazekileyo ngamanye amaxesha zaziphucula kwimisebenzi yeevidiyo yomxholo omde. Ukusasazeka apha akusiyorhafu yomgangatho oyihlawulelayo isantya; kwizinto ezingenayo ezinde kubonakala kusimahla.

Kutheni ufanele ukuyikhathalela njengomthengi: ixabiso lokuhambisa lilo elimisela ixabiso loluhlu. Isiseko esinexabiso lengqwalasela eliyeka ukukhula sisiseko onokusifumana ngexabiso ukuze usalathise kwimifanekiso elikhulu lamawaka. Yiyo leyo indlela esebenza ngayo "eyona modeli yombono engacinezelwanga itshiphu" — asilomlinganiselo wokukhuthaza, sisakhiwo.

04Ithetha ntoni i-3BIT-REAP, kwaye ikubiza phi

Imodeli yeeparamitha ezingamakhulu amane anamashumi amabini anesibhozo ezigidigidi ayilingani nayiphi na indawo elula kukuchaneka okupheleleyo. Isakhiwo sethu se-MiniMax-M3 sithwala impumezelo ye-3BIT-REAP, echaza imisebenzi emibini eyahlukeneyo esetyenziswe ngolandelelwano oluthile.

REAP: cima iingcali phambi kokuba uzicinezele

I-REAP — Router-weighted Expert Activation Pruning — yindlela yokucinezela i-MoE yesitshotsho esinye evela kwi-Cerebras naseYunivesithi yaseCalgary, epapashwe njenge-arXiv 2510.13999 neyamkelwa kwi-ICLR 2026. Ingxoxo yayo yeyokuba kwimisebenzi yokuvelisa, ukuthena iingcali kubetha ukuzidibanisa: ukudibanisa kutshabalalisa ulawulo olucokisekileyo lwerawuta yaye kwazisa impazamo ongasoze ukwazi ukuyibuyisa. Ngoko i-REAP endaweni yoko inika inqaku ingcali nganye ngomlinganiselo wokubaluleka odibanisa amaxabiso egeyithi yerawuta (ukuba irawuta iyikhetha kangaphi nangamandla angakanani) kunye nemilinganiselo yokusebenza kweengcali (ukuba imveliso yayo ifaka isandla ngokwenene kangakanani), ize icime ngokuthe ngqo iingcali ezibaluleke kancinci.

Ababhali baxela ukususa ukuya kuma-50% eengcali kwiimodeli ezinkulu njengeziyi-1 trillion yeeparamitha ngoxa begcina umgangatho osisiseko — kwi-Qwen3-480B-Coder-FP8 kukuthena kwe-50%, ama-97.6% esakhono sokukhowuda esingesiso se-agentic esisisiseko nama-96.7% kwi-SWE-Bench ye-agentic agcinwa. Lawo ngamanani ephepha kwiimodeli zephepha, kungekhona awethu kwi-MiniMax-M3; siyawacaphula ukuze ukwazi ukuthelekelela indlela, kungekhona ukubonisa ukuba kukho ulinganiselo olulinganayo apha.

Ubunzima beebhithi ezintathu, kwesibini

Ubunzima obusindileyo emva koko bugcinwa kwiibhithi ezintathu. Ulandelelwano lwenziwe ngabom kwaye yiyo injongo yolu dibaniso: thena kuqala, ucuthe kwesibini. Akukho ngqiqo ekuchitheni iibhithi zakho ezintathu ezixabisekileyo kubunzima ngabunye ngokuthembekileyo ukhowuda iingcali osele uza kuzicima. Thena umsila osasazekileyo, uze uvumele uhlahlo-lwabiwo-mali oluseleyo luye ngokupheleleyo kwiingcali ezidubulayo ngokwenene.

Isilumkiso esinyanisekileyo

I-3BIT-REAP linqanaba elinamandla kunazo zonke esilihambisayo. Silisebenzisa kwezona ziseko zinkulu — i-MiniMax-M3, i-Kimi-K3, i-DeepSeek-V4-Pro, i-GLM-5.2, i-Nemotron3Ultra — kuba ngumahluko phakathi kokuba ezo modeli zifumaneke ngexabiso elisengqiqweni nokungafumaneki kwaphela. Apho isiseko sincinci ngokwaneleyo ukuba singasidingi, asisisebenzisi: iimodeli ze-W8A16 kwikhathalogu ziqhuba ubunzima obuyi-8-bit ngokusebenza okuyi-16-bit, kwaye i-W4A16-AUTOROUND-REAP ihlala phakathi, isebenzisa ukujikeleza okufundiweyo kwe-AutoRound ukubuyisa ukuchaneka okulahlwa kukucuthwa okujikelezela kweyona ikufuphi ngokungenabuchule. Khetha inqanaba elihambelana nonyamezelo lwakho. Ukuba ukhupha izinto kumatyala eshumi lamawaka, iibhithi ezintathu yintengiselwano echanekileyo. Ukuba wenza ukufunda okunye kwezoqhaqho okubaluleke kakhulu komfanekiso omnye, chitha imali kwi-Kimi-K3.

05Kuthetha ntoni "akukho maleko wokwala" xa into engenayo ingumfanekiso

Ukungacinezelwa ligama elisetyenziswa ngokungakhathali, ngoko nali ibango elichanekileyo: akukho maleko wokwala kwaye akukho kuhluzwa kwemveliso kule endpoint. Akukho sahluli-ndidi sihlola umfanekiso wakho phambi kokuba imodeli iwubone. Akukho modeli yomgaqo-nkqubo ihlola iitoken zisiya ngaphandle. Into eveliswa yimodeli yiyo oyifumanayo.

Kumbhalo, umphumo waloo nto uyaziwa. Kumbono, wahlukile ngendidi, kuba ukuziphatha kokwala kwimifanekiso engenayo kulapho ii-API zorhwebo zeendlela ezininzi ziqinile khona kwaye zenza ngokungahleliweyo. Ukusilela okuqhelekileyo kubonakala ngolu hlobo:

• Ukwala okupheleleyo kwesazisi — ukwala ukuchaza abantu efotweni kwaphela, kubandakanya iimpawu ezibonakala ngokucacileyo nezingenangozi ngokupheleleyo, kuba umgaqo-nkqubo awukwazi ukwahlula inkcazelo kwisazisi.
• Ukuphuncuka ngenxa yomxholo — ukwala ukubhala okanye ukushwankathela uxwebhu ngenxa yoko luthetha ngako, nangona ukubhala kungumsebenzi wemitshini.
• Ukoyika kwezonyango nakwezoqhaqho — ukwala ukuchaza imifanekiso yeklinikhi, amaxwebhu okwenzakala, okanye iifoto zeengozi umsebenzi ozizingcali afuna zichazwe.
• Iimpazamo zezixhobo zokhuseleko — ukwala ukufunda umfanekiso wesikrini kuba unetheminali, umbamba wemiqulu, okanye isilungisi.
• Ukwehla okuthe cwaka — eyona imbi: hayi ukwala kodwa impendulo engeyompendulo, edlula kuqinisekiso lwesikimu sakho ize itshabalalise iseti yakho yedatha ngaphandle kokuvusa impazamo.

Ingxaki yokusebenza kwezi zintlanu ayisiyoyokuba azinabulungisa. Yeyokuba azinakulinganiswa. Impendulo engachanekanga linqaku ledatha; unokulinika inqaku, ulibambe ekuvavanyweni, uze ulungise ngenxa yalo. Ukwala ngumngxuma. Kubungakanani beqela, umlinganiselo wokwala we-3% osasazwe ngokungahleliswanga kwizinto zakho ezingenayo — ogxile, ngokwakhiwa, kanye kumatyala anzima nangaqhelekanga — ntyekelo yenkqubo kwiziphumo zakho engabuyiswayo naliphi na inani lococeko lwangasemva.

I-Shannon iqhutywa yi-Shannon Lab LLC, New Mexico, USA, phantsi kwePolisi yethu yoSetyenziso oluNoxanduva. Ukungacinezelwa kuthetha ukuba imodeli ayigwebi ngezinto zakho ezingenayo; akuthethi ukuba i-akhawunti ayinamigaqo.

06Iinkcukacha ezipheleleyo

Indawo	MiniMax-M3 · 3BIT-REAP
Isazisi semodeli	MiniMax-M3-3BIT-REAP
Isiseko	MiniMax-M3 (MiniMax AI)
Ifestile yomxholo	iitoken ezingama-262,144
Imveliso ephezulu	iitoken ezingama-16,384
Umfanekiso ongenayo	Ewe
Ividiyo engenayo	Ayivezwanga
Ukusasaza	Ewe
Ukubiza izixhobo	Ewe
Ukuqiqa	Ewe (reasoning_effort)
Imveliso ecwangcisiweyo	json_schema + response_format
Umaleko wokwala	Akukho
Ii-endpoint	/v1/chat/completions · /v1/messages · /v1/responses
Ihanjiswa kwi-	iqoqo leeGPU le-Shannon

Imigca emibini kuloo theyibhile ifanele ukugxininiswa. Ividiyo engenayo ayivezwanga. Isiseko se-M3 sangaphambili saqeqeshwa kwiividiyo yaye i-MiniMax ibhala ngeevidiyo ezingenayo; i-endpoint yethu yamkela imifanekiso. Yakha ngemifanekiso. Okwesibini, ifestile yeetoken ezingama-262,144 yeyethu, ingelilo inani leetoken ezisigidi lasekuqaleni — cwangcisa imisebenzi yamaxwebhu amade neyemifanekiso emininzi ngokwe-262,144 kwaye awusoze umangaliswe.

07Ukubiza i-API ye-MiniMax engacinezelwanga

Iilwimi ezintathu, imodeli enye emva kwazo zonke, ukusasaza kuzo zontathu. Sebenzisa nayiphi na esele ithethwa ngumthengi wakho okhoyo; akukho mahluko wesakhono phakathi kwazo. Yalathisa i-URL yakho esisiseko kwi-Shannon uze utshintshe isazisi semodeli.

Ukuqonda umfanekiso ngolwimi lwe-OpenAI

curl https://api.shannon-ai.com/v1/chat/completions \
  -H "Authorization: Bearer $SHANNON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-M3-3BIT-REAP",
    "messages": [{
      "role": "user",
      "content": [
        { "type": "text", "text": "Transcribe every field in this scan. Do not summarize." },
        { "type": "image_url",
          "image_url": { "url": "data:image/png;base64,iVBORw0KGgo..." } }
      ]
    }],
    "max_tokens": 4096,
    "stream": true
  }'

Ii-URL zakude ze-https:// nee-URL zedatha ye-base64 zombini ziyasebenza kwinxalenye ye-image_url. Ncamathelisa iinxalenye zemifanekiso ezininzi kumyalezo omnye xa udinga uthelekiso lwamaphepha — unamawaka angama-262 anekhulu elinamashumi amane anesine eetoken zokuchitha kwaye, nge-$0.50 kwitoken enye yezigidi, ukuzichitha kutshiphu.

Ukukhupha okunodidi ngesikimu se-JSON

Apho le modeli izifumanela khona indawo yayo kumgca wokusebenza kudibaniso lombono + isikimu esingqongqo + ixabiso elisezantsi. Ayizizo zonke iimodeli kwikhathalogu yethu ezinazo zontathu: i-Laguna-S-2.1 ne-inkling azixhasi i-json_schema. I-MiniMax-M3 iyayixhasa, nto leyo ithetha ukuba unokufumana into eqinisekisiweyo kumfanekiso ngaphandle komaleko wokwahlula:

{
  "model": "MiniMax-M3-3BIT-REAP",
  "messages": [ /* text part + image_url part */ ],
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "invoice",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "vendor":     { "type": "string" },
          "issued_on":  { "type": "string" },
          "total_cents":{ "type": "integer" },
          "line_items": { "type": "array", "items": { "type": "string" } }
        },
        "required": ["vendor", "issued_on", "total_cents", "line_items"],
        "additionalProperties": false
      }
    }
  }
}

Amandla okuqiqa

Imodeli ivula i-reasoning_effort. Kwiindidi zemifanekiso ngobuninzi, yigcine isezantsi — ubuza umbuzo wokubona, kwaye iitoken zokuqiqa ziitoken eziphumayo ozihlawulelayo. Yinyuse kwimibizo apho umfanekiso ubungqina kwingxoxo ende: ukufunda umzobo ngokuchasene neenkcukacha, ukuvumelanisa itshati netheyibhile, ukugqiba ukuba ingaba imifanekiso yesikrini emibini ibonisa ukungalungi okufanayo na.

Ireferensi epheleleyo yeeparamitha, ukuchazwa kweempazamo, kunye neefomathi zokusasaza kuzo zontathu iilwimi kuhlala kwi-amaxwebhu e-API ye-Shannon.

08Apho i-MiniMax-M3 ilukhetho olululo — naphi apho ingelulo

Yifikelele xa

• Umsebenzi wemifanekiso ngobuninzi. Ukubhalwa kweeleyibhile zeeseti zedatha, uphando lolawulo lomxholo, ukutyebisa ikhathalogu, ukucubungula imifanekiso yesikrini, i-OCR nesakhiwo ngeqela kumaxwebhu askeniweyo. Ixabiso elisezantsi lingxoxo iphela.
• Imifanekiso ibe zizingxelo ezinodidi. Umbono kunye ne-json_schema engqongqo kunye neetoken ezitshiphu kunazo zonke kwikhathalogu ludibaniso olunikwa ziimodeli ezimbini kuphela apha.
• Umxholo omde weendlela ezininzi. Amaphepha amaninzi, okanye izikrini ezininzi, kwisicelo esinye — isakhiwo esingaphantsi sakhelwe kanye lo milo wezinto ezingenayo.
• Uphando apho ukwala konakalisa isampuli. Naluphi na uphononongo apho ubunyani buxhomekeke kwimodeli ephendula into nganye, kubandakanya nezo zingakhululekisiyo.
• Imijikelo ye-agentic efunda isikrini. Umnombo woluhlu lwe-M lu-agentic ngokucacileyo, kwaye ukubiza izixhobo kuyafumaneka.

Fikelela kwenye into xa

• Ufuna uphahla lombono oluphezulu kunazo zonke kwaye uhlahlo-lwabiwo-mali lukhona. I-Kimi-K3 · 3BIT-REAP iphindwe izihlandlo eziyi-7.7 kwixabiso lezingenayo ngesizathu; kukufunda okunye okubaluleke kakhulu, leyo yimali echanekileyo ekufuneka uyichithile.
• Ufuna ubunzima obuthembeke kakhulu. Iibhithi ezintathu linqanaba lethu elinamandla kunazo zonke. Iimodeli ze-W8A16 zitshintshisa ixabiso ngokuchaneka, kwaye i-MiMo-V2.5 · W8A16 inombono kwixabiso elinye le-$0.50 / $2.00 ukuba ukuthembeka kubunzima ngabunye kuloo nto uyilungiselelayo.
• Ufuna ividiyo. Asiyivezi ividiyo engenayo kule endpoint.
• Umsebenzi ngumbhalo kuphela kumphambili. I-DeepSeek-V4-Pro-0813 ikhona ngenxa yoko; ubuya kuhlawulela inqaba yombono ongasoze uyisebenzise.

Amanani esakhiwo, obalo-mgangatho nawokukhawuleza osapho lwe-MiniMax axelwe yilebhu ye-MiniMax AI; amanani e-REAP axelwe ngababhali be-REAP kwiimodeli zabo yaye ayilulo ulinganiselo lwesi sakhiwo. Amaxabiso e-Shannon, ifestile yomxholo, neeflegi zesakhono zezethu kwaye zifundwa kwikhathalogu yethu ephilayo yeemodeli.

09Imibuzo ebuzwa rhoqo

---

Imithombo: Isikhokelo seModeli ye-MiniMax-M3 · MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention · MiniMax-01: Scaling Foundation Models with Lightning Attention · REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression. Amanani obalo-mgangatho nawokukhawuleza e-MiniMax axelwe yilebhu kwaye asexeshana kulindwe uqinisekiso oluzimeleyo. Amaxabiso e-Shannon neeflegi zesakhono ahamba ngokwe-5 Septemba 2026.