API yeMiniMax isina kudzorwa: kuona pamutengo wepasi
MiniMax-M3 ndiyo nzira yakachipa kupfuura dzose yekuwana kunzwisisa mifananidzo kusina kudzorwa paGedhi reShannon — $0.50 anopinda, $2.00 anobuda, matokeni 262,144 econtext, uye hapana chidzitiro chekuramba pakati pako nemufananidzo.
Pfupiso
MiniMax-M3 · 3BIT-REAP imodhi inoona, isina kudzorwa, iri pamutengo wepasi weShannon: $0.50 pamatokeni miriyoni anopinda ne$2.00 pamatokeni miriyoni anobuda, ine context yematokeni 262,144, matokeni 16,384 anobuda, kuyerera, kudana zvishandiso, kufungisisa kwaine reasoning_effort, nezvinobuda zvakasimba zvejson_schema. Yakavakwa pahwaro hweM3 hweMiniMax AI — modhi yeMixture-of-Experts ine mapharamita 428B ine MiniMax Sparse Attention nekudzidziswa kwemidziyo yakawanda kwepachayo — yakadzvanywa nehuremu hwe3-bit nekubvisa nyanzvi kweREAP kuti ikwane pacluster yedu yemaGPU pamutengo iwoyo. Kana basa rako riri rekuti "tarisa mufananidzo uyu upindure nechokwadi, zvuru gumi," ndiyo modhi yacho.
Anenge API yega yega inogamuchira midziyo yakawanda inoisa kuona semutengo wepamusoro. Unobhadhara chiyero chemodhi yemavara pamavara, uye wozobhadhara zvakawanda — dzimwe nguva zvakapetwa kanoverengeka — panguva chaiyo iyo mufananidzo unopinda muprompt. Pamusoro peizvozvo, mifananidzo inopinda ndipo pane maitiro ekuramba akaomarara kupfuura pese: modhi inofara kukurukura nyaya iri mumavara icharamba kutsanangura chinhu chimwe chetecho chiri mumufananidzo. Chinobuda ndechigadzirwa chinodhura pakudana kwega kwega uye chisingavimbike pakudana kwega kwega, ndicho chinhu chakaipa kupfuura zvose kana uchigadzirisa huwandu hukuru. MiniMax-M3 paShannon ndiyo mhinduro yedu pane izvozvo: $0.50 / $2.00 imwe cheteyo yaunobhadhara pamodhi yedu yemavara yakachipa kupfuura dzose, ichishandiswa pahwaro hwakadzidziswa nemidziyo yakawanda kubva padanho rekutanga, pasina chidzitiro chekuramba chakamira mberi kwahwo.
01Kuti "modhi inoona isina kudzorwa yakachipa kupfuura dzose" inorevei chaizvo pano
Katarogu yeShannon imamodhi gumi nemaviri. Mashanu acho anogamuchira mifananidzo. Mitengo yawo, pamatokeni miriyoni:
| Modhi inoona | Anopinda | Anobuda | Zvinobuda zvakarongwa |
|---|---|---|---|
| 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 chete |
| Kimi-K3 · 3BIT-REAP | $3.83 | $19.12 | json_schema + response_format |
MiniMax-M3 iri pasi chaipo — $0.50 / $2.00 — yakaenzana neMiMo-V2.5 uye hapana chiri pasi payo, nokuti $0.50 / $2.00 ndiwo mutengo wepasi werondedzero yedu. Kuenzaniswa nepamusoro pemamodhi anoona, yakachipa ka7.7 pane zvinopinda ne ka9.6 pane zvinobuda kupfuura Kimi-K3.
Isa izvozvo pabasa chairo. Ngatiti uri kurongedza mifananidzo 1,000 — masikirinishoti, mapeji akaskenwa, mifananidzo yezvigadzirwa — ine avheji yematokeni anenge 1,200 emufananidzo pamwe neprompt achipinda nematokeni 300 emhinduro achibuda:
Mitengo yerondedzero yeShannon, Gunyana 2026. Huwandu hwematokeni ndehwekuratidzira; kuverengwa kwematokeni emifananidzo yako kunotsamira pachiyero chemufananidzo nekuti mifananidzo mingani yaunonamira pakudana kwega kwega.
Zvimwe zviviri zvinobatsira kuchengeta mumusoro. Kuzadza context yose yematokeni 262,144 kamwe chete kunodhura anenge masendi gumi nematatu. Mhinduro yakareba pamusoro-soro yematokeni 16,384 inodhura anenge masendi matatu nehafu. Ndiyo miganhu chaiyo yechikumbiro chimwe chakaipa kupfuura zvose, uye idiki zvekuti chinodzora hurongwa hwekuona hwako chinorega kuva mari chichitanga kuva kubudirira kwehuwandu.
02Dzinza reMiniMax: rabhoritari inogadzirisa matokeni, kwete misoro yenhau
MiniMax AI irabhoritari yeShanghai ine nhoroondo yose yehuremu hwakavhurika iri pamubvunzo mumwe chete: unoita sei kuti attention ichipe zvekuti context refu nemidziyo yakawanda zvirege kuva zvinhu zvevapfumi? Kubudiswa kwega kwega mumutsara uyu imhinduro yemubvunzo iwoyo, uye mhinduro dzinovakirana.
MiniMax-01 — Lightning Attention, uye kuona kubva pakutanga
Mutsara weMiniMax-01 (wakabudiswa searXiv 2501.08313) wakaunza Lightning Attention, kuita kunoziva I/O kwerudzi rweattention yakatwasuka, mukati memodhi yeMixture-of-Experts yakasanganiswa ine mapharamita 456B ose ne45.9B anoshanda patokeni imwe neimwe. Mutsara wakauya sepeya: MiniMax-Text-01 neMiniMax-VL-01, modhi inoona. VL-01 yakashandisa chimiro cheViT-MLP-LLM chinozivikanwa, asi ine Vision Transformer ine mapharamita mamiriyoni 303 yakadzidziswa kubva pakutanga pamapeya emifananidzo netsananguro mamiriyoni 694 pane muisi wakakwereterwa, ipapo yakarongeswa nekugadziriswa pamwe chete kuburikidza nehurongwa hwematanho akawanda hwakagadzirisa matokeni mabhiriyoni 512.
Chinhu ichocho chinokosha kuchinyorwa chaunoverenga. Kuona mudzinza rino hakusi chinhu chakazonamirwa pamodhi yemavara kuti chikwanise kuratidzwa. MiniMax yave ichidzidzisa shongwe yayo yekuona kubva pakubudiswa kwayo kwekutanga kwakavhurika, ndicho chikonzero kunzwisisa magwaro, kunzwisisa maUI nemasikirinishoti, nekuverenga machati zvave zviri simba rakasimba remhuri iyi pane zvinhu zvakazofungwa.
MiniMax-M1 — attention yakasanganiswa yekufungisisa kwenguva refu
MiniMax-M1 (arXiv 2506.13585) yakatsanangurwa nevanyori vayo semodhi yekutanga ine huremu hwakavhurika, huru, inofungisisa ine attention yakasanganiswa. Yakachengeta muviri weMoE we456B / 45.9B ndokurukanidza marudzi maviri eattention: bhuroko retransformer rine softmax attention yakajairika mushure mebhuroko rega rega rechinomwe reLightning Attention. Mubairo wakayerwa mukushanda pane pachinzvimbo cherondedzero — MiniMax yakashuma kuti pakureba kwekugadzira kwematokeni 100K, M1 yakashandisa anenge 25% eFLOPs aDeepSeek R1, apo ichitsigira context yematokeni miriyoni imwe nenzira yepachayo.
MiniMax-M2 — kutendeukira kumaactivation madiki
MiniMax-M2 yakaenda kune rumwe rutivi pakukura ikawana kuremekedzwa nekuda kwezvo: dhikodha yemadanho 62 ine mapharamita 229.9B ose asi ane ~9.8B chete anoshanda patokeni imwe neimwe, ichishandisa nyanzvi diki-diki 256 padanho rimwe nerimwe reMoE ne routing yesigmoid kusvika kusere padziri, nzvimbo dzeRoPE neQK-RMSNorm dzekugarisisa kwekudzidziswa. Yaiva pachena modhi yekukodha yemaagent — kugadziridza mafaira akawanda, madenderedzwa ekumhanyisa nekugadzirisa, kugadzirisa kunoongororwa nemiedzo — uye yakaunza chimiro chekufungisisa kwakarukanidzwa, apo kufungisisa kunotakurwa pakati pemadenderedzwa ekudana zvishandiso panzvimbo pekuraswa mushure mekutenderera kwega kwega. M2 ndiyo modhi yakavaka mukurumbira weMiniMax wekupa maitiro emaagent ari pedyo nepamberi pachikamu chidiki chemutengo patokeni, nekungochengeta maactivation ari madiki.
MiniMax-M3 — hwaro huri pasi pesuwo rino
MiniMax-M3 yakaburitswa musi wa1 Chikumi 2026, nehuremu paHugging Face nokukurumidza mushure meizvozvo negwaro reunyanzvi paarXiv 2606.13392. Imodhi yeMixture-of-Experts ine mapharamita 428B ine anenge 23B anoshanda patokeni imwe neimwe, context yakanyorwa yematokeni miriyoni imwe, nemifananidzo nemavhidhiyo anopinda epachayo — MiniMax inoitsanangura seyakadzidziswa nemidziyo yakasanganiswa "kubva padanho rekutanga chairo," pane kudzidziswa pamavara ipapo yozogadziridzwa kuti igamuchire midziyo yakawanda. Panhamba dzayo dzakabudiswa inowana 80.5% paSWE-bench Verified ne59.0% paSWE-bench Pro. Idzo inhamba dzakashumwa nemutengesi pamuyero wekukodha, ndiyo chaiyo nzira yaunofanira kudzitora nayo; tinodzitaura nokuti ndizvo zvakataurwa nerabhoritari, kwete nokuti takazviyera pachedu.
03MiniMax Sparse Attention: sei modhi inoona ichigona kuchipa kudai
Unyanzvi hunonakidza muM3 ndiMiniMax Sparse Attention (MSA), uye ndicho chikonzero chakananga chekuti modhi yemidziyo yakawanda yehukuru uhwu inogona kugara pamutengo wedu wepasi. Mifananidzo inodya matokeni. Mapeji mashoma echiyero chepamusoro anogona kuisa matokeni zviuru makumi mumuprompt usati wanyora murairo mumwe chete. Pasi peattention yakajairika, mutengo iwoyo unopeta kaviri, uye modhi inoverenga mapeji makumi maviri haidhuri kaviri kupfuura inoverenga gumi — inodhura zvakanyanya kupfuura kaviri.
MSA inochinja chimiro chegwara iroro. Ichimiro chine mabhuroko akapararira chakavakwa pahwaro hweGrouped Query Attention, chakakamurwa kuita mativi maviri:
- Rutivi rweIndex rwakareruka runopa mabhuroko ekiyi-nekukosha zviyero — rwuchitora zviyero zvepamusoro pamatokeni ari mubhuroko rimwe nerimwe — ropasarudza top-k panzvimbo yega yega yemubvunzo nepaboka rimwe nerimwe reGQA, ruchigara ruchibatanidza bhuroko remuno rine mubvunzo pachawo kuti kudzidziswa kugarisise.
- Rutivi Rukuru runomhanyisa softmax attention yakajairika pamabhuroko akasarudzwa neindex chete.
Kurongwa kuri mugwaro reMSA ndik = mabhuroko 16 ematokeni 128, izvo zvinopa mubvunzo wega wega bhajeti rakatarwa rematokeni 2,048 anotariswa. Rakatarwa, zvichireva kuti harikuri nekureba kwemutsara. Ndiwo mano ose: sezvo context ichikura, mutengo weattention wehwaro hwakazara unoramba uchikwira apo mutengo werutivi rukuru rweMSA unogara wakati sandara, saka attention inokura zvakatwasuka pane kupeta kaviri.
Nhamba dzinobva mugwaro reMiniMax Sparse Attention (arXiv 2606.13392): kudzikira kwe28.4× kwemaFLOPs eattention patokeni kuenzaniswa neGQA yakarongwa zvakafanana pacontext ye1M, nekumhanya kwe14.2× paprefill / 7.6× padecode kwakayerwa panhumbi dzeH800 muongororo yegwaro yemapharamita 109B. MiniMax inoshuma zvakasiyana kumhanya kwe9× paprefill ne15× padecode kwaM3 kuenzaniswa neM2 pacontext ye1M. Zvose zvakashumwa nerabhoritari.
Gwaro rinoitawo ongororo inokosha kuchinyorwa chekuona: paMoE ye109B inogamuchira midziyo yakawanda nenzira yepachayo, MSA yakachengeta kushanda kwakaenzana nehwaro hweattention yakazara pamiyero yemifananidzo inosanganisira AI2D, ChartQA neMMMU, nepamiyero yemavhidhiyo inosanganisira EgoSchema, LongVideoBench neVideoMME — apo marudzi akapararira dzimwe nguva achivandudza pamabasa emavhidhiyo econtext refu. Kupararira pano hakusi mutero wemhando waunobhadhara nekuda kwekumhanya; pazvinopinda zvakareba zvinoita sekunge kwemahara.
Sei izvi zvichifanira kukubata semutengi: mutengo wekushandisa ndiwo unotara mutengo werondedzero. Hwaro hune mutengo weattention usingakuri ndihwo hwaunogona kukwanisa kunongedzera pamifananidzo zviuru zana. Ndiwo mano ari mumashure me"modhi inoona isina kudzorwa yakachipa kupfuura dzose" — hausi mutengo wekunyengedza, imagadzirirwo.
04Kuti 3BIT-REAP inorevei, uye pairinokudhurira
Modhi ine mapharamita mabhiriyoni mazana mana nemakumi maviri nemasere haikwani panzvimbo iri nyore pakururama kuzere. Chivakwa chedu cheMiniMax-M3 chine chipfupiso che3BIT-REAP, chinotsanangura mabasa maviri akaparadzana akaitwa nenzira yakatarwa.
REAP: bvisa nyanzvi usati wadzidzvanya
REAP — Router-weighted Expert Activation Pruning — inzira yekudzvanya MoE kamwe chete inobva kuCerebras neUniversity of Calgary, yakabudiswa searXiv 2510.13999 uye yakagamuchirwa kuICLR 2026. Nharo yayo ndeyekuti pamabasa ekugadzira, kubvisa nyanzvi kunokunda kudzisanganisa: kusanganisa kunoparadza kudzora kwakatsetseka kwerouter uye kunounza chikanganiso chausingazogoni kudzorera. Saka panzvimbo peizvozvo REAP inopa nyanzvi imwe neimwe chiyero chinosanganisa zviyero zvegedhi rerouter (kuti router inoisarudza kangani uye zvakasimba zvakadii) nezviyero zvekushanda kwenyanzvi (kuti zvinobuda zvayo zvinobatsira zvakadii), ipapo yobvisa nyanzvi dzisingakoshi zvachose.
Vanyori vanoshuma kubvisa kusvika ku50% yenyanzvi pamamodhi makuru anosvika tiriyoni imwe yemapharamita ichichengetedza mhando yepakutanga — paQwen3-480B-Coder-FP8 pakubvisa kwe50%, zvakachengetwa zvaiva 97.6% yeunyanzvi hwekukodha husiri hwemaagent ne96.7% paSWE-Bench yemaagent. Idzo inhamba dzegwaro pamamodhi egwaro, kwete dzedu paMiniMax-M3; tinodzitaura kuti utonge nzira yacho, kwete kuti tireve kuti pane kuyera kwakaenzana pano.
Huremu hwe3-bit, zvichitevera
Huremu hwasara hunozochengetwa pamabhiti matatu. Kutevedzana kwacho kwakaitwa nemaune uye ndicho chinangwa chemusanganiswa: bvisa kutanga, quantize zvichitevera. Hazvina musoro kushandisa mabhiti ako matatu anokosha pahuremu humwe nehumwe uchinyatsochengeta nyanzvi dzaunenge uchida kubvisa. Bvisa muswe wakapararira, wozorega bhajeti rasara richienda rose kunyanzvi dzinonyatsoshanda.
Chinoyambirwa chechokwadi
3BIT-REAP ndiyo mhando yakaomarara kupfuura dzose yatinoshandisa. Tinoishandisa pahwaro hukuru kupfuura hwose — MiniMax-M3, Kimi-K3, DeepSeek-V4-Pro, GLM-5.2, Nemotron3Ultra — nokuti ndiwo musiyano uri pakati pemamodhi iwayo kuwanikwa pamutengo une musoro nekusawanikwa zvachose. Apo hwaro hudiki zvekuti hahuidi, hatiishandisi: mamodhi eW8A16 ari mukatarogu anomhanya nehuremu hwe8-bit nemaactivation e16-bit, uye W4A16-AUTOROUND-REAP iri pakati, ichishandisa kutenderedza kwakadzidzwa kweAutoRound kudzorera kururama kunorasikirwa nekutenderedza kwakapusa kune iri pedyo. Sarudza mhando inoenderana nekukwanisa kwako. Kana uchibvisa minda kubva pamainvoyisi zviuru gumi, 3-bit ndiko kuchinjana kwakakodzera. Kana uchiita kuverenga kumwe chete kweforensic kunokosha kwemufananidzo mumwe, shandisa mari paKimi-K3.
05Kuti "hapana chidzitiro chekuramba" inorevei kana chinopinda chiri mufananidzo
Kusadzorwa ishoko rinoshandiswa zvisina kunyatsotarwa, saka heino nharo chaiyo: hapana chidzitiro chekuramba uye hapana sefa yezvirimo pane zvinobuda pasuwo rino. Hapana muongorori anotarisa mufananidzo wako modhi isati yauona. Hapana modhi yemutemo inotarisa matokeni panzira yekubuda. Zvinogadzirwa nemodhi ndizvo zvaunogamuchira.
Mumavara, chinobva ipapo chinozivikanwa. Pakuona, chakasiyana zvachose, nokuti maitiro ekuramba pamifananidzo inopinda ndipo pane maAPI ekutengeserana emidziyo yakawanda achinyanya kuchenjerera uye achinyanya kusanzwisisika. Kutadza kwacho kunoita seizvi:
- Kuramba kwakazara pamusoro pekuzivikanwa — kuramba kutsanangura vanhu vari mumufananidzo zvachose, kusanganisira zvinhu zvinoonekwa pachena uye zvisina njodzi zvachose, nokuti mutemo haugoni kusiyanisa kutsanangura nekuzivisa kuti ndiani.
- Kutiza nyaya — kuramba kunyorazve kana kupfupisa gwaro nekuda kwenyaya iri mugwaro, kunyange kunyorazve kuri basa remuchina.
- Kutya kwezvekiriniki neforensic — kuramba kutsanangura mifananidzo yekiriniki, zvinyorwa zvekukuvara, kana mifananidzo yetsaona inodiwa nemunyanzvi kuti itsanangurwe.
- Kuramba kusina musoro pazvishandiso zvekuchengetedza — kuramba kuverenga sikirinishoti nokuti ine terminal, packet capture, kana debugger.
- Kudzikira kusinganzwiki — chakaipa kupfuura zvose: kwete kuramba asi mhinduro isina kujeka, inopfuura pakuongororwa kweschema yako uye ichikuvadza dhata rako pasina kumbobudisa chikanganiso.
Dambudziko rekushanda pane zvose zvishanu harisi rekuti hazvina kunaka. Nderekuti hazviyerwe. Mhinduro isiriyo ipoinzi yedhata; unogona kuipa chiyero, kuibata pakuongorora, nekuigadzirisa. Kuramba ndiwo mwena. Pahuwandu hukuru, chiyero chekuramba che3% chakapararira zvisiri zvechisero pazvinopinda zvako — chakaunganidzwa, nemagadzirirwo, pane zvinhu zvakaoma uye zvisina kujairika chaizvo — kutsauka kwakarongwa mumhedzisiro yako kusingadzoswi nekuchenesa kupi zvako kwemashure. Kubvisa chidzitiro chekuramba kunobvisa chinounza dhata rinoshaikwa nenzira yakanyarara, yakabatana.
Shannon inoshandiswa neShannon Lab LLC, New Mexico, USA, pasi peMutemo wedu Wekushandisa Nemutsigo. Kusadzorwa kunoreva kuti modhi haitongi zvinopinda zvako; hakurevi kuti akaundi haina mitemo.
06Zvirongwa zvakazara
| Munda | MiniMax-M3 · 3BIT-REAP |
|---|---|
| Zita remodhi | MiniMax-M3-3BIT-REAP |
| Hwaro | MiniMax-M3 (MiniMax AI) |
| Hwindo recontext | matokeni 262,144 |
| Zvinobuda zvakanyanya | matokeni 16,384 |
| Mifananidzo inopinda | Hongu |
| Vhidhiyo inopinda | Haiwanikwi |
| Kuyerera | Hongu |
| Kudana zvishandiso | Hongu |
| Kufungisisa | Hongu (reasoning_effort) |
| Zvinobuda zvakarongwa | json_schema + response_format |
| Chidzitiro chekuramba | Hapana |
| Masuwo | /v1/chat/completions · /v1/messages · /v1/responses |
| Inoshandiswa pa | cluster yedu yemaGPU |
Mitsara miviri iri mutafura iyoyo inoda kusimbiswa. Vhidhiyo inopinda haiwanikwi. Hwaro hweM3 hwakadzidziswa pamavhidhiyo uye MiniMax inonyora nezvevhidhiyo inopinda; suwo redu rinogamuchira mifananidzo. Vaka uchishandisa mifananidzo. Chechipiri, hwindo rematokeni 262,144 nderedu, kwete nhamba yemiriyoni imwe yekumusoro — ronga mabasa emagwaro marefu neemifananidzo yakawanda uchitarisa 262,144 uye hauzombokatyamadzwi.
07Kudana API yeMiniMax isina kudzorwa
Mitauro mitatu, modhi imwe cheteyo iri mumashure mayo yose, kuyerera pamitatu yose. Shandisa iyo mutengi wako wazvino anotoziva; hapana musiyano weunyanzvi pakati padzo. Nongedzera base URL yako kuShannon uchinje zita remodhi.
Kunzwisisa mifananidzo kuburikidza nemutauro weOpenAI
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
}'MaURL ekure ehttps:// nemaURL edhata ebase64 ose anoshanda muchikamu cheimage_url. Namira zvikamu zvemifananidzo zvakawanda mumeseji imwe chete kana uchida kuenzanisa mapeji — une matokeni 262,144 ekushandisa uye, pa$0.50 pamiriyoni, kuashandisa kwakachipa.
Kubvisa ruzivo rune mhando neJSON schema
Apo modhi iyi inowana nzvimbo yayo muhurongwa ndipo pamusanganiswa wekuona + schema yakasimba + mutengo wepasi. Haasi mamodhi ose mukatarogu yedu ane zvitatu zvose: Laguna-S-2.1 neinkling hazvitsigiri json_schema. MiniMax-M3 inotsigira, zvinoreva kuti unogona kuwana chinhu chakaongororwa kubva pamufananidzo pasina danho rekuparadzanisa:
{
"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
}
}
}
}Simba rekufungisisa
Modhi inoburitsa reasoning_effort. Pakurongedza mifananidzo yakawanda, richengete rakadzikira — uri kubvunza mubvunzo wekuona, uye matokeni ekufungisisa matokeni anobuda aunobhadhara. Rikwidze pakudana apo mufananidzo uri humbowo munharo dzakareba: kuverenga dhayagiramu uchienzanisa nezvirongwa, kubatanidza chati netafura, kusarudza kana masikirinishoti maviri achiratidza chikanganiso chimwe chete.
Rondedzero yakazara yemapharamita, zvinorehwa nezvikanganiso, nezvimiro zvekuyerera zvemitauro mitatu yose zviri mumagwaro eAPI eShannon.
08Apo MiniMax-M3 iri sarudzo yakakodzera — nepaisiri
Isarudze kana
- Basa remifananidzo rehuwandu hukuru. Kunyora marebhuru pamadhata, tsvakurudzo yekudzora zvirimo, kuwedzera ruzivo pakatarogu, kurongedza masikirinishoti, OCR-nekurongedza kwehuwandu pamagwaro akaskenwa. Mutengo wepasi ndiyo nharo yose.
- Mifananidzo ichiitwa zvinyorwa zvine mhando. Kuona pamwe ne
json_schemayakasimba pamwe nematokeni akachipa kupfuura ose mukatarogu musanganiswa unopiwa nemamodhi maviri chete pano. - Context yakareba yemidziyo yakawanda. Mapeji akawanda, kana mafuremu akawanda, muchikumbiro chimwe chete — magadzirirwo ari pasi apa akavakirwa chimiro chezvinopinda ichi chaicho.
- Tsvakurudzo apo kuramba kunokuvadza mienzaniso. Ongororo ipi zvayo ine chokwadi chinotsamira pakuti modhi ipindure chinhu chega chega, kusanganisira zvinonetsa.
- Madenderedzwa emaagent anoverenga sikirini. Dzinza reM-series nderemaagent zviri pachena, uye kudana zvishandiso kunowanikwa.
Sarudza chimwe kana
- Uchida denga repamusoro rekuona uye mari iripo. Kimi-K3 · 3BIT-REAP inodhura ka7.7 pane zvinopinda nechikonzero; pakuverenga kumwe chete kunokosha, ndiyo mari yakakodzera kushandisa.
- Uchida huremu hune kutendeseka kwepamusoro. 3-bit ndiyo mhando yedu yakaomarara kupfuura dzose. Mamodhi e
W8A16anochinjanisa mutengo nekururama, uye MiMo-V2.5 · W8A16 inoona pamutengo mumwe chete we$0.50 / $2.00 kana kutendeseka pahuremu humwe nehumwe ndicho chaunotsvaga. - Uchida vhidhiyo. Hatiburitsi vhidhiyo inopinda pasuwo rino.
- Basa nderemavara chete padanho repamberi. DeepSeek-V4-Pro-0813 iripo nekuda kweizvozvo; waizobhadhara shongwe yekuona yausingatombobati.
Nhamba yega yega yekunze iri muchinyorwa chino yakabudiswa nemumwe munhu asiri isu. Tarisa zvakatorwa pane kutora pfupiso yedu yazvo.
Nhamba dzemagadzirirwo, dzemiyero nedzekumhanya dzemhuri yeMiniMax dzakashumwa nerabhoritari yeMiniMax AI; nhamba dzeREAP dzakashumwa nevanyori veREAP pamamodhi avo uye hadzisi kuyera kwechivakwa ichi. Mitengo yeShannon, hwindo recontext nezviratidzo zveunyanzvi ndezvedu uye zvinoverengwa kubva pakatarogu yedu yemamodhi iri kushanda.
09Mibvunzo inowanzobvunzwa
Chii chinonzi API yeMiniMax isina kudzorwa?
NdiMiniMax-M3 · 3BIT-REAP inoshandiswa pacluster yedu yemaGPU pasina chidzitiro chekuramba uye pasina sefa yezvirimo pane zvinobuda. Unoidana kuburikidza ne/v1/chat/completions, /v1/messages, kana /v1/responses ine hwindo recontext rematokeni 262,144, mifananidzo inopinda, kudana zvishandiso, nezvinobuda zvakarongwa zveJSON-schema, pa$0.50 pamatokeni miriyoni anopinda ne$2.00 pamatokeni miriyoni anobuda.
MiniMax-M3 paShannon inogamuchira mifananidzo inopinda here?
Hongu. MiniMax-M3 ndeimwe yemamodhi mashanu anoona ari mukatarogu yeShannon uye inogamuchira mifananidzo muchimiro chakajairika chechikamu chezvirimo cheimage_url, kusanganisira maURL edhata ebase64. Suwo redu rinoburitsa mifananidzo inopinda chete — hatiburitsi vhidhiyo inopinda, kunyange hwaro hwekumusoro hwakadzidziswa pamavhidhiyo.
Kunzwisisa mifananidzo kusina kudzorwa kunodhura marii?
MiniMax-M3 iri pamutengo wepasi weShannon we$0.50 pamatokeni miriyoni anopinda ne$2.00 pamatokeni miriyoni anobuda. Boka rekudana kwemifananidzo 1,000 kune avheji yematokeni 1,200 anopinda nematokeni 300 anobuda pakudana kwega kwega rinodhura anenge $1.20. Boka rimwe chetero paKimi-K3, modhi inoona ine denga repamusoro mukatarogu, rinodhura anenge $10.34.
3BIT-REAP inorevei, uye inokuvadza mhando here?
REAP iRouter-weighted Expert Activation Pruning, nzira yekudzvanya Mixture-of-Experts kamwe chete inobva kuCerebras neUniversity of Calgary inopa nyanzvi imwe neimwe chiyero maererano nekukosha kwegedhi rerouter nechiyero chekushanda kwayo, ipapo yobvisa nyanzvi dzisingakoshi quantization isati yaitwa. Huremu hwasara hunochengetwa pamabhiti matatu. Ndiyo mhando yedu yekudzvanya yakaomarara kupfuura dzose uye ndiyo inoita kuti hwaro hwemapharamita mabhiriyoni mazana mana nemakumi maviri nemasere hukwanise kushandiswa pamutengo wepasi. Kana uchida kutendeseka kwepamusoro-soro tinopa mamodhi akaparadzana eW8A16 panzvimbo peizvozvo.
Inotsigira kudana zvishandiso nezvinobuda zveJSON schema here?
Hongu pazviri zviviri. MiniMax-M3 inotsigira kuyerera, kudana zvishandiso, kufungisisa kwaine chidzoreso chereasoning_effort, response_format, nezvinobuda zvakarongwa zvakasimba zvejson_schema. Musanganiswa iwoyo hausi pamamodhi ose ari mukatarogu — Laguna-S-2.1 neinkling hazvitsigiri json_schema — saka kana uchida chinhu chine mhando chichidzoka kubva pamufananidzo, MiniMax-M3 ndiyo modhi yakachipa kupfuura dzose ichakupa chimwe.
Ndinowana hwindo recontext nemuganhu wezvinobuda upi?
Matokeni 262,144 econtext nematokeni 16,384 anobuda pachikumbiro chimwe nechimwe, zvakafanana pamamodhi gumi nemaviri ose eShannon. Hwaro hweMiniMax-M3 hwekumusoro hwakanyorwa pamatokeni miriyoni imwe; isu tinopa hwindo re262K, saka ronga mabasa emagwaro marefu neemifananidzo yakawanda uchitarisa 262,144, kwete nhamba yekumusoro.
Edza API yeMiniMax isina kudzorwa
Kunzwisisa mifananidzo pamutengo wepasi. Mitauro mitatu, zita remodhi rimwe chete, hapana chidzitiro chekuramba.
Verenga Magwaro eAPI Nhungamiro Dzemamodhi OseShannon Lab LLC · New Mexico, USA · Mutemo Wekushandisa Nemutsigo unoshanda
Zvakatorwa kubva: kadhi remodhi reMiniMax-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. Nhamba dzemiyero nedzekumhanya dzeMiniMax dzakashumwa nerabhoritari uye dziri panguva yekumirira kuongororwa kwakazvimirira. Mitengo yeShannon nezviratidzo zveunyanzvi zviripo kusvika musi wa5 Gunyana 2026.