I-API ye-MiniMax engavinjelwe: umbono ngentengo ephansi

Cishe zonke ama-API ezindlela eziningi asingathiwe abeka umbono njengento ekhethekile ebizayo. Ukhokha izinga lemodeli yombhalo ngombhalo, bese ukhokha ngaphezulu — kwesinye isikhathi ngokuphindwe kaningi — ngokushesha nje uma isithombe singena emyalelweni. Ngaphezu kwalokho, okufakwayo okuyisithombe yilapho ukuziphatha kokwenqaba kuqine khona kakhulu: imodeli ezoxoxa ngenjabulo ngesihloko esithile embhalweni izokwenqaba ukuchaza into efanayo esithombeni. Umphumela wumkhiqizo obiza kakhulu ekubizeni ngakunye futhi ongathembeki ekubizeni ngakunye, okuyikho okubi kunakho konke uma ucubungula ivolumu. I-MiniMax-M3 ku-Shannon yimpendulo yethu kulokho: u-$0.50 / $2.00 ofanayo owukhokhelayo emodelini yethu yombhalo eshibhe kunazo zonke, usetshenziswa esisekelweni esaqeqeshwa ngezindlela eziningi kusukela esinyathelweni sokuqala, ngaphandle kongqimba lokwenqaba oluhlezi phambi kwaso.

01Kusho ukuthini ngempela u-"imodeli yombono engavinjelwe eshibhe kunazo zonke" lapha

Ikhathalogu ye-Shannon ingamamodeli ayishumi nambili. Amahlanu kuwo amukela izithombe. Amanani awo, ngethokheni eyisigidi:

Imodeli enombono	Okungenayo	Okuphumayo	Okukhishwayo okuhlelekile
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 phansi kakhulu — $0.50 / $2.00 — ilingana ne-MiMo-V2.5 futhi ayikho engaphansi kwayo, ngoba u-$0.50 / $2.00 uyilo iphansi lohlu lwethu lwamanani. Uma iqhathaniswa nophahla lwebanga lombono ishibhe ngokuphindwe ka-7.7 kokungenayo nangokuphindwe ka-9.6 kokuphumayo kune-Kimi-K3.

Beka lokho emsebenzini. Ake sithi uhlunga izithombe eziyi-1,000 — izithombe-skrini, amakhasi askeniwe, izithombe zomkhiqizo — okulinganiselwa emathokhenini angu-1,200 esithombe kanye nomyalelo angena namathokheni angu-300 empendulo aphuma:

Amanani ohlu lwe-Shannon, Septhemba 2026. Izibalo zamathokheni ziyisibonelo; ukuhlukaniswa kwesithombe sakho kube ngamathokheni kuxhomeke kunani lamaphikseli nakwinani lezithombe ozinamathisela ekubizeni ngakunye.

Amanye amaphuzu amabili ereferensi kuwusizo ukuwagcina ekhanda. Ukugcwalisa wonke umongo wamathokheni angu-262,144 kanye kubiza cishe amasenti angu-13. Impendulo yobude obuphezulu bamathokheni angu-16,384 ibiza cishe amasenti angu-3.3. Leyo yimikhawulo yangempela yesicelo esisodwa esibi kunazo zonke, futhi mincane ngokwanele ukuze umkhawulo othakazelisayo kumshini wombono uyeke ukuba yimali uqale ukuba ngukukhishwa.

02Uzalo lwe-MiniMax: ilebhu elungiselela amathokheni, hhayi izihloko zamaphephandaba

I-MiniMax AI yilebhu yase-Shanghai umlando wayo wonke wezisindo ezivulekile omayelana nombuzo owodwa: ungakwenza kanjani ukunaka kushibhe ngokwanele ukuze umongo omude nezindlela eziningi kuyeke ukuba yinto yokunethezeka? Ukukhishwa ngakunye kulolu hlu kuyimpendulo kulowo mbuzo, futhi izimpendulo zakhelana.

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

Uchungechunge lwe-MiniMax-01 (olwashicilelwa njenge-arXiv 2501.08313) lwethula i-Lightning Attention, ukusetshenziswa okuqaphela i-I/O kwenguqulo yokunaka okuqondile, ngaphakathi kwemodeli exubile ye-Mixture-of-Experts yamapharamitha angu-456B esewonke nangu-45.9B asebenzayo kuthokheni ngalinye. Uchungechunge lwathunyelwa njengepheya: i-MiniMax-Text-01 kanye ne-MiniMax-VL-01, imodeli yombono. I-VL-01 yasebenzisa isitaki esesijwayelekile se-ViT-MLP-LLM, kodwa nge-Vision Transformer yamapharamitha ayizigidi ezingu-303 eqeqeshwe kusukela ekuqaleni emapheyeni ezithombe-namagama ayizigidi ezingu-694 kunokusebenzisa i-encoder eboleked, yase iqondaniswa futhi ilungiswa ndawonye emshinini onezigaba eziningi ocubungula amathokheni ayizigidigidi ezingu-512.

Leyo mininingwane ibalulekile kulesi sihloko osifundayo. Umbono kulolu zalo awusiyo into eyanamathiselwa emuva emodelini yombhalo ukuze kufakwe uphawu ebhokisini. I-MiniMax ibilokhu iqeqesha umbhoshongo wayo wokubona kusukela ekukhishweni kwayo kokuqala okuvulekile, yingakho ukuqonda amadokhumenti, ukuqonda ubuxhakaxhaka besikrini nezithombe-skrini, kanye nokufunda amashadi kube ngamandla aqhubekayo omndeni kunokuba yizinto ezicatshangwe ngemuva.

MiniMax-M1 — ukunaka okuxubile ukuze kucatshangwe isikhathi eside

I-MiniMax-M1 (arXiv 2506.13585) yachazwa ngababhali bayo njengemodeli yokuqala yokucabanga yezisindo ezivulekile enokunaka okuxubile ngesilinganiso esikhulu. Yagcina umzimba we-MoE ongu-456B / 45.9B futhi yaxuba izinhlobo ezimbili zokunaka: ibhlogo ye-transformer ye-softmax-attention ejwayelekile ngemuva kwamabhlogo ayisikhombisa e-Lightning Attention. Umvuzo wakalwa ngokubala kunendawo yebhodi lezikhundla — i-MiniMax ibike ukuthi ebudeni bokukhiqiza obungamathokheni angu-100K, i-M1 yasebenzisa cishe u-25% wama-FLOP e-DeepSeek R1, kuyilapho isekela ngokwemvelo umongo wamathokheni ayisigidi.

MiniMax-M2 — ukuphenduka kokusebenza okuncane

I-MiniMax-M2 yaya ngenye indlela ngosayizi futhi yazuza ukunakwa ngenxa yalokho: i-decoder yamangqimba angu-62 ene-mapharamitha angu-229.9B esewonke kodwa angu-9.8B kuphela asebenzayo kuthokheni ngalinye, isebenzisa ochwepheshe abaqondiswayo abancane abangu-256 engqimbeni ngalinye le-MoE ngokuqondiswa kwe-sigmoid kwabayisishiyagalombili kubo, izindawo ze-RoPE kanye ne-QK-RMSNorm ukuze ukuqeqeshwa kuzinze. Yayiyimodeli yekhodi ye-ejenti ngokusobala — ukuhlela amafayela amaningi, amaluphu okusebenza-nokulungisa, ukulungiswa okuqinisekiswe ngokuhlolwa — futhi yethula ifomethi yokucabanga okuxubile, lapho ukucabanga kuthwalwa phakathi kwemizuliswano yokubiza amathuluzi kunokulahlwa ngemuva komzuliswano ngamunye. I-M2 yiyona modeli eyakha idumela le-MiniMax lokuletha ukuziphatha kwe-ejenti eliseduze nomkhawulo ngengxenye encane yezindleko ngethokheni ngalinye, ngokugcina ukusebenza kumncane nje.

MiniMax-M3 — isisekelo esingaphansi kwalesi siphetho

I-MiniMax-M3 yakhishwa ngoNhlangulana 1, 2026, ngezisindo ku-Hugging Face ngemuva kwesikhashana kanye nombiko wobuchwepheshe ku-arXiv 2606.13392. Iyimodeli ye-Mixture-of-Experts yamapharamitha angu-428B enamapharamitha acishe abe ngu-23B asebenzayo kuthokheni ngalinye, inomongo wamathokheni ayisigidi obhaliwe, kanye nokufakwa kwezithombe namavidiyo kwemvelo — i-MiniMax iyichaza njengeqeqeshwe ngezindlela ezixubile "from the very first step," kunokuqeqeshwa ngombhalo kuqala bese kwengezwa izindlela eziningi. Emananini ayo ashicilelwe i-MiniMax ithola u-80.5% ku-SWE-bench Verified kanye no-59.0% ku-SWE-bench Pro. Lawo ngamanani abikwe ngumthengisi ekuhlolweni kwekhodi, okuyiyona ndlela ncamashi okufanele uwakale ngayo; siwacaphuna ngoba ayizimangalo ezishicilelwe zelebhu, hhayi ngoba sike sawakhiqiza kabusha ngokuzimela.

03I-MiniMax Sparse Attention: kungani imodeli yombono ingaba shibhe kangaka

Ubunjiniyela obuthakazelisayo ku-M3 yi-MiniMax Sparse Attention (MSA), futhi yiso ngqo isizathu sokuthi imodeli yezindlela eziningi yalolu sayizi ingakwazi ukuhlala entengweni yethu ephansi kunazo zonke. Izithombe zibiza amathokheni amaningi. Amakhasi ambalwa anamaphikseli amaningi angafaka amashumi ezinkulungwane zamathokheni emyalelweni ngaphambi kokuba ubhale umyalelo owodwa. Ngaphansi kokunaka okujwayelekile, lezo zindleko zanda ngokuphindwe kabili, futhi imodeli efunda amakhasi angamashumi amabili ayibizi ngokuphindwe kabili kuneyifunda ayishumi — kubi kakhulu kunokuphindwe kabili.

I-MSA ishintsha isimo salowo mugqa. Iyindlela ye-block-sparse eyakhelwe phezu kwesisekelo se-Grouped Query Attention, ehlukaniswe kabili:

• I-Index Branch elula enikeza amabhlogo ekhiye-nanani amaphuzu — inikeza amaphuzu e-max-pooling emathokhenini angaphakathi kwebhlogo ngalinye — bese ikhetha i-top-k endaweni ngayinye yombuzo naseqenjini ngalinye le-GQA, ihlala ifaka ibhlogo lendawo eliqukethe umbuzo uqobo ukuze ukuqeqeshwa kuzinze.
• I-Main Branch eqhuba ukunaka okujwayelekile kwe-softmax emabhlogweni akhethwe yi-index kuphela.

Ukumiswa ephepheni le-MSA ngu-k = 16 amabhlogo amathokheni angu-128, okunika umbuzo ngamunye isabelo esingaguquki samathokheni angu-2,048 anakiwe. Esingaguquki, okusho ukuthi asikhuli nobude bochungechunge. Yilelo iqhinga lonke: njengoba umongo ukhula, izindleko zokunaka zesisekelo esiminyene ziyaqhubeka zenyuka kanti izindleko ze-main branch ye-MSA zihlala zilingana, ngakho ukunaka kukhula ngokuqondile kunokuphindwa kabili.

Amanani asuka embikweni we-MiniMax Sparse Attention (arXiv 2606.13392): ukwehla okuphindwe ka-28.4 kwama-FLOP okunaka ngethokheni ngalinye uma kuqhathaniswa ne-GQA emiswe ngokufanayo emongweni we-1M, kanye nokusheshiswa kwewashi kwe-prefill okuphindwe ka-14.2 nokwe-decode okuphindwe ka-7.6 okwakalwa kuma-hardware e-H800 ocwaningweni lwephepha lwamapharamitha angu-109B. I-MiniMax ngokwehlukene ibika ukusheshiswa kwe-prefill okuphindwe ka-9 nokwe-decode okuphindwe ka-15 kwe-M3 uma kuqhathaniswa ne-M2 emongweni we-1M. Konke kubikwe yilebhu.

Iphepha liphinde lenze ukuhlola okubalulekile esihlokweni sombono: ku-MoE engu-109B yezindlela eziningi zemvelo, i-MSA yagcina ukusebenza okulinganayo nesisekelo sokunaka okugcwele kuwo wonke ukuhlolwa kwezithombe okufaka i-AI2D, i-ChartQA ne-MMMU, nakuwo wonke ukuhlolwa kwamavidiyo okufaka i-EgoSchema, i-LongVideoBench ne-VideoMME — kanti izinhlobo ezisabalele ngezinye izikhathi zathuthukisa emisebenzini yamavidiyo enomongo omude. Ukusabalala lapha akuyona intela yekhwalithi oyikhokhelayo isivinini; kokufakwayo okude kubonakala cishe kumahhala.

Kungani kufanele ukhathazeke njengekhasimende: izindleko zokukhonza yizo ezibeka intengo yohlu. Isisekelo esiyeka ukukhula ngezindleko zokunaka yisisekelo ongakwazi ukusikhomba ezithombeni eziyizinkulungwane eziyikhulu. Yilowo umshini ongemuva kokuthi "imodeli yombono engavinjelwe eshibhe kunazo zonke" — akulona izinga lokukhangisa, kungukwakhiwa.

04Kusho ukuthini u-3BIT-REAP, nokuthi kukubiza kuphi

Imodeli yamapharamitha ayizigidigidi ezingu-428 ayilingani ndawo enethezekile ngokunemba okugcwele. Ukwakhiwa kwethu kwe-MiniMax-M3 kuphethe isijobelelo esithi 3BIT-REAP, esichaza imisebenzi emibili ehlukene esetshenziswe ngokulandelana okuthile.

I-REAP: susa ochwepheshe ngaphambi kokubacindezela

I-REAP — Router-weighted Expert Activation Pruning — yindlela yokucindezela i-MoE yokushaya kanye evela e-Cerebras nase-University of Calgary, eyashicilelwa njenge-arXiv 2510.13999 futhi yamukelwa ku-ICLR 2026. Impikiswano yayo ukuthi emisebenzini yokukhiqiza, ukuthena ochwepheshe kudlula ukubahlanganisa: ukuhlanganisa kubhubhisa ukulawula okuncane kwe-router futhi kwethula iphutha ongeke ulibuyise. Ngakho i-REAP kunalokho inikeza uchwepheshe ngamunye amaphuzu ngesisekelo sokubaluleka esihlanganisa amanani egeti ye-router (ukuthi kangaki futhi ngamandla amangakanani i-router imkhetha) namazinga okusebenza kochwepheshe (ukuthi okukhishwayo kwakhe kufaka kangakanani ngempela), bese ikhipha ngokuphelele ochwepheshe abangabalulekile kunabo bonke.

Ababhali babika ukususa kufika ku-50% wochwepheshe emamodelini amakhulu njengalawo anamapharamitha ayisigidintathu kuyilapho begcina ikhwalithi eyisisekelo — ku-Qwen3-480B-Coder-FP8 ekuthenweni kwama-50%, kwagcinwa u-97.6% wekhono lekhodi eliyisisekelo elingelona ele-ejenti no-96.7% ku-SWE-Bench ye-ejenti. Lawo ngamanani ephepha emamodelini ephepha, hhayi awethu ku-MiniMax-M3; siwacaphuna ukuze wahlulele indlela, hhayi ukusho ukuthi kukhona ukukala okulinganayo lapha.

Izisindo zamabhithi amathathu, kwesibili

Izisindo ezisalayo bese zigcinwa emabhithini amathathu. Ukulandelana kwenziwa ngamabomu futhi kuyiphuzu lokuhlanganiswa: thena kuqala, ukale kwesibili. Akunangqondo ukuchitha amabhithi akho amathathu ayigugu esisindweni ngasinye ubhala ngokwethembeka ochwepheshe osuzobasusa. Thena umsila osabalele, bese uvumela isabelo esisele siye ngokuphelele kochwepheshe abasebenza ngempela.

Isixwayiso esiqotho

U-3BIT-REAP uyizinga eliqine kunawo wonke esiwathumelayo. Siwusebenzisa ezisekelweni ezinkulu kunazo zonke — i-MiniMax-M3, i-Kimi-K3, i-DeepSeek-V4-Pro, i-GLM-5.2, i-Nemotron3Ultra — ngoba yiwo umehluko phakathi kwalawo mamodeli etholakala ngentengo enengqondo nokungatholakali nhlobo. Lapho isisekelo sisincane ngokwanele ukuze singawudingi, asiwusebenzisi: amamodeli e-W8A16 ekhathalogu asebenzisa izisindo zamabhithi ayisishiyagalombili nokusebenza kwamabhithi ayi-16, kanti u-W4A16-AUTOROUND-REAP uhlala phakathi, usebenzisa ukugoba okufundiwe kwe-AutoRound ukubuyisa ukunemba okulahlwa ngukukalwa okulula kwe-round-to-nearest. Khetha izinga elihambisana nokubekezela kwakho. Uma ukhipha izinkambu kuma-invoyisi ayizinkulungwane eziyishumi, amabhithi amathathu ngukushintshana okulungile. Uma wenza ukufundwa okukodwa kwe-forensic okubalulekile kwesithombe esisodwa, chitha imali ku-Kimi-K3.

05Kusho ukuthini u-"alukho ungqimba lokwenqaba" uma okufakwayo kuyisithombe

Okungavinjelwe yigama elisetshenziswa ngokuxegayo, ngakho nasi isimangalo esiqondile: alukho ungqimba lokwenqaba futhi akukho ukuhlungwa kokukhishwayo kulesi siphetho. Asikho isihlukanisi esihlola isithombe sakho ngaphambi kokuba imodeli isibone. Ayikho imodeli yenqubomgomo ehlola amathokheni endleleni aphuma ngayo. Lokho imodeli ekukhiqizayo yikho okutholayo.

Embhalweni, umphumela walokho uyajwayeleka. Embonweni, kwehlukile ngohlobo, ngoba ukuziphatha kokwenqaba kokufakwayo okuyisithombe yilapho ama-API ezentengiso ezindlela eziningi eqine khona kakhulu futhi engacacile khona kakhulu. Ukwehluleka okuvamile kubukeka kanje:

• Ukwenqaba okuphelele kobunikazi — ukwenqaba ukuchaza abantu esithombeni nhlobo, kufaka nezici ezibonakala ngokusobala futhi ezingenabungozi nhlobo, ngoba inqubomgomo ayikwazi ukwehlukanisa incazelo nokukhomba umuntu.
• Ukubaleka ngenxa yesihloko — ukwenqaba ukubhala noma ukufingqa idokhumenti ngenxa yalokho idokhumenti eyimayelana nakho, nakuba ukubhala kungumsebenzi wemishini.
• Ukwesaba kwezokwelapha nokwe-forensic — ukwenqaba ukuchaza izithombe zomtholampilo, ukubhalwa kokulimala, noma izithombe zezingozi umuntu ochwepheshile adinga ukuthi zichazwe.
• Ukwenqaba kwamanga kwamathuluzi okuphepha — ukwenqaba ukufunda isithombe-skrini ngoba siqukethe i-terminal, ukubanjwa kwephakethe, noma i-debugger.
• Ukwehla okuthule — okubi kunakho konke: hhayi ukwenqaba kodwa impendulo evikelayo engaphenduli, edlula ekuqinisekisweni kwesikimu sakho futhi ifake ubuthi eqoqweni lakho ledatha ngaphandle kokuveza iphutha.

Inkinga yokusebenza kuzo zonke lezi ezinhlanu akusikho ukuthi azinabo ubulungiswa. Ukuthi azikaleki. Impendulo engalungile iyiphuzu ledatha; ungayinika amaphuzu, uyibambe ekuhlolweni, futhi uyilungise. Ukwenqaba kuyimbobo. Ngesilinganiso seqoqo, izinga lokwenqaba elingu-3% elisatshalaliswe ngokungahleliwe kokufakwayo kwakho — okugxile, ngokwakhiwa, ngqo ezimweni ezinzima nezingavamile — kuwukuchema okuhlelekile emiphumeleni yakho okungabuyiswa yikho konke ukuhlanzwa kwangemuva. Ukususa ungqimba lokwenqaba kususa umthombo wedatha elahlekile ethule futhi ehlobene.

I-Shannon iqhutshwa yi-Shannon Lab LLC, New Mexico, USA, ngaphansi kwenqubomgomo yethu yokusebenzisa ngokuzibophezela. Okungavinjelwe kusho ukuthi imodeli ayiniki umbono ngokufakwayo kwakho; akusho ukuthi i-akhawunti ayinamigomo.

06Umqulu ophelele wezakhi

Inkambu	MiniMax-M3 · 3BIT-REAP
I-id yemodeli	MiniMax-M3-3BIT-REAP
Isisekelo	MiniMax-M3 (MiniMax AI)
Iwindi lomongo	amathokheni angu-262,144
Okukhishwayo okuphezulu	amathokheni angu-16,384
Okufakwayo okuyisithombe	Yebo
Okufakwayo okuyividiyo	Akuvezwa
Ukusakaza	Yebo
Ukubiza amathuluzi	Yebo
Ukucabanga	Yebo (reasoning_effort)
Okukhishwayo okuhlelekile	json_schema + response_format
Ungqimba lokwenqaba	Alukho
Iziphetho	/v1/chat/completions · /v1/messages · /v1/responses
Ikhonzwa ku-	iqoqo lama-GPU le-Shannon

Imigqa emibili kulelo thebula ifanele ukugcizelelwa. Okufakwayo okuyividiyo akuvezwa. Isisekelo se-M3 esingenhla saqeqeshwa ngamavidiyo futhi i-MiniMax ibhala okufakwayo okuyividiyo; isiphetho sethu samukela izithombe. Yakhela izithombe. Okwesibili, iwindi lamathokheni angu-262,144 ngelethu, hhayi inani lezigidi elingaphezulu — hlelela imisebenzi yamadokhumenti amade nezithombe eziningi ngokumelene no-262,144 futhi ngeke uke umangale.

07Ukubiza i-API ye-MiniMax engavinjelwe

Izilimi ezintathu, imodeli efanayo ngemuva kwazo zonke, ukusakaza kuzo zontathu. Sebenzisa noma yiluphi iklayenti lakho elikhona oselulukhuluma; awukho umehluko wamakhono phakathi kwazo. Khomba i-URL yakho eyisisekelo ku-Shannon bese ushintsha i-id yemodeli.

Ukuqonda izithombe ngolimi 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
  }'

Ama-URL akude e-https:// nama-URL edatha e-base64 womabili ayasebenza engxenyeni ye-image_url. Namathisela izingxenye ezimbalwa zesithombe emlayezweni owodwa uma udinga ukuqhathanisa amakhasi — unamathokheni angu-262,144 ongawachitha futhi, ku-$0.50 ngesigidi, ukuwachitha kushibhile.

Ukukhipha ulwazi olunohlobo nge-JSON schema

Lapho le modeli izuza khona indawo yayo emshinini yikuhlanganiswa kombono + isikimu esiqinile + intengo ephansi kunazo zonke. Akuwona wonke amamodeli ekhathalogu yethu anakho kokuthathu: i-Laguna-S-2.1 ne-inkling azimseki u-json_schema. I-MiniMax-M3 iyamseka, okusho ukuthi ungathola into eqinisekisiwe esithombeni ngaphandle kongqimba lokuhlaziya:

{
  "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
      }
    }
  }
}

Umzamo wokucabanga

Imodeli iveza u-reasoning_effort. Ekuhlukaniseni izithombe ngobuningi, wugcine uphansi — ubuza umbuzo wokubona, futhi amathokheni okucabanga angamathokheni okukhishwayo owakhokhelayo. Wukhuphule ekubizeni lapho isithombe siwubufakazi empikiswaneni ende: ukufunda umdwebo ngokumelene nemininingwane, ukuhlanganisa ishadi nethebula, ukunquma ukuthi izithombe-skrini ezimbili zikhombisa iphutha elifanayo yini.

Ireferensi ephelele yamapharamitha, incazelo yamaphutha, kanye namafomethi okusakaza kuzo zontathu izilimi kutholakala emibhalweni ye-API ye-Shannon.

08Lapho i-MiniMax-M3 iyisinqumo esifanele — nalapho ingesona

Yifinyelele uma

• Umsebenzi wezithombe ovolumu ephezulu. Ukumaka amaqoqo edatha, ucwaningo lokulawula okuqukethwe, ukucebisa ikhathalogu, ukuhlunga izithombe-skrini, i-OCR yeqoqo nokuhlelwa kwamadokhumenti askeniwe. Intengo ephansi kunazo zonke yiyo yonke impikiswano.
• Izithombe ziba ngamarekhodi anohlobo. Umbono kanye no-json_schema oqinile kanye namathokheni ashibhe kunawo wonke ekhathalogu kungukuhlanganiswa amamodeli amabili kuphela lapha akunikezayo.
• Umongo omude wezindlela eziningi. Amakhasi amaningi, noma amafreyimu amaningi, esicelweni esisodwa — ukwakhiwa okungaphansi kwakhelwe ngqo lolu hlobo lokufakwayo.
• Ucwaningo lapho ukwenqaba konakalisa isampula. Noma yiluphi ucwaningo ubuqiniso balo obuxhomeke ekutheni imodeli iphendule yonke into, kufaka nezingakhululi.
• Amaluphu e-ejenti afunda isikrini. Uzalo lochungechunge lwe-M lungolwe-ejenti ngokusobala, futhi ukubiza amathuluzi kuyatholakala.

Yifinyelele kokunye uma

• Udinga uphahla lombono oluphezulu kunazo zonke futhi isabelomali sikhona. I-Kimi-K3 · 3BIT-REAP iphindwe ka-7.7 ngentengo yokungenayo ngesizathu; ekufundeni okukodwa okubalulekile, leyo yimali efanele ukuchithwa.
• Udinga izisindo ezethembeke kakhulu. Amabhithi amathathu ayizinga lethu eliqine kunawo wonke. Amamodeli e-W8A16 ashintshanisa intengo ngokunemba, kanti i-MiMo-V2.5 · W8A16 inombono ngentengo efanayo ka-$0.50 / $2.00 uma ukwethembeka esisindweni ngasinye kuyikho okulungiselelayo.
• Udinga ividiyo. Asikuvezi okufakwayo okuyividiyo kulesi siphetho.
• Umsebenzi ungumbhalo ompoqompoqo osemkhawulweni. I-DeepSeek-V4-Pro-0813 ikhona ngenxa yalokho; ubuzobe ukhokhela umbhoshongo wombono ongawusebenzisi.

Amanani okwakhiwa, okuhlolwa nokusheshiswa omndeni we-MiniMax abikwa yilebhu ye-MiniMax AI; amanani e-REAP abikwa ngababhali be-REAP emamodelini abo futhi akusikho ukukalwa kwalokhu kwakhiwa. Amanani e-Shannon, iwindi lomongo namafulegi amakhono ngawethu futhi afundwa ekhathalogu yethu yamamodeli ephilayo.

09Imibuzo ebuzwa njalo

---

Imithombo: Ikhadi lemodeli le-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 okuhlolwa nokusheshiswa e-MiniMax abikwa yilebhu futhi asalindile ukuqinisekiswa okuzimele. Amanani e-Shannon namafulegi amakhono angawamanje kusukela ngoMandulo 5, 2026.