API yosasidwa ya MiniMax: masomphenya pa mtengo wa pansi

Pafupifupi API iliyonse ya multimodal yotumikiridwa imalipiritsa masomphenya ngati chinthu chapamwamba. Mumalipira mtengo wa mtundu wa mawu pa mawu, kenako mumalipira zambiri — nthawi zina kuwirikiza kangapo — chithunzi chikangolowa mu funso. Kuwonjezera pa izo, zolowetsa za zithunzi ndiko kumene khalidwe la kukana limakhala lamphamvu kwambiri: mtundu umene ungakambirane mutu mosangalala mu mawu udzakana kufotokoza chinthu chomwecho mu chithunzi. Zotsatira zake ndi chinthu chokwera mtengo pa yitanidwe ndi chosadalirika pa yitanidwe, chimene ndi choipa kwambiri pa mbali zonse ziwiri pamene mukukonza kuchuluka. MiniMax-M3 pa Shannon ndi yankho lathu pa izo: $0.50 / $2.00 omwewo amene mumalipira pa mtundu wathu wa mawu wotsika mtengo kwambiri, ogwiritsidwa ntchito pa maziko amene anaphunzitsidwa mwa multimodal kuyambira sitepe yoyamba, popanda wosanjikiza wa kukana patsogolo pake.

01Chomwe "mtundu wa masomphenya wosasidwa wotsika mtengo kwambiri" chikutanthauza pano

Katalogi ya Shannon ndi mitundu khumi ndi iwiri. Isanu mwa iyo imalandira zithunzi. Mitengo yawo, pa miliyoni ya ma token:

Mtundu wokhoza masomphenya	Kulowa	Kutuluka	Zotuluka zokonzedwa
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 yokha
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	json_schema + response_format

MiniMax-M3 ili pa pansi — $0.50 / $2.00 — yofanana ndi MiMo-V2.5 ndipo palibe chotsika kuposa pamenepo, chifukwa $0.50 / $2.00 ndiye pansi pa mndandanda wathu wa mitengo. Motsutsana ndi pamwamba pa gulu la masomphenya ndi yotsika kuwirikiza 7.7× pa kulowa ndi kuwirikiza 9.6× pa kutuluka kuposa Kimi-K3.

Ikani izo mu ntchito. Tiyerekeze kuti mukusanthula zithunzi 1,000 — zithunzi za skrini, masamba osindikizidwa, zithunzi za zinthu — ndi pafupifupi ma token 1,200 a chithunzi kuphatikiza funso kulowa ndi ma token 300 a yankho kutuluka:

Mitengo ya mndandanda ya Shannon, Seputembala 2026. Ziwerengero za ma token ndi zosonyeza chabe; kuwerengedwa kwa ma token a chithunzi chanu kumadalira resolution ndi kuchuluka kwa zithunzi zomwe mumamatira pa yitanidwe lililonse.

Mfundo zina ziwiri zoyenera kuzikumbukira. Kudzaza context yonse ya ma token 262,144 kamodzi kumalipiritsa pafupifupi masenti 13. Yankho la kutalika kwakukulu la ma token 16,384 limalipiritsa pafupifupi masenti 3.3. Amenewo ndi malire enieni a pempho limodzi loipa kwambiri, ndipo ndi ang'onoang'ono mokwanira kuti chotchinga chosangalatsa pa njira ya masomphenya chisiye kukhala ndalama ndipo chikhale kuchuluka kwa ntchito.

02Mzere wa MiniMax: labotale imene imakonza ma token, osati mitu ya nkhani

MiniMax AI ndi labotale ya ku Shanghai imene mbiri yake yonse ya zolemera zotseguka ikukhudza funso limodzi: mumapanga bwanji chisamalira kukhala chotsika mtengo mokwanira kuti context yaitali ndi multimodality zisiye kukhala zinthu zapamwamba? Kutulutsidwa kulikonse mu mzerewu ndi yankho la funso limenelo, ndipo mayankho amamangirira wina ndi mnzake.

MiniMax-01 — Lightning Attention, ndi masomphenya kuyambira pachiyambi

Mndandanda wa MiniMax-01 (wofalitsidwa ngati arXiv 2501.08313) unayambitsa Lightning Attention, kukhazikitsidwa kodziwa I/O kwa mtundu wa linear attention, mkati mwa mtundu wosakanikirana wa Mixture-of-Experts wa zigawo zonse 456B ndi 45.9B zogwira ntchito pa token iliyonse. Mndandandawu unatulutsidwa ngati awiri: MiniMax-Text-01 ndi MiniMax-VL-01, mtundu wa masomphenya. VL-01 unagwiritsa ntchito chulu cha ViT-MLP-LLM chodziwika tsopano, koma ndi Vision Transformer ya zigawo mamiliyoni 303 yophunzitsidwa kuyambira pachiyambi pa awiriawiri a chithunzi ndi mawu mamiliyoni 694 m'malo mwa encoder yobwereka, kenako yogwirizanitsidwa ndi kukonzedwa pamodzi mu njira ya magawo ambiri yokonza ma token mabiliyoni 512.

Tsatanetsatane umenewo ndi wofunika pa nkhani imene mukuwerenga. Masomphenya mu mzerewu si adaputala womangiriridwa mochedwa pa mtundu wa mawu kuti awonjezere chinthu. MiniMax yakhala ikuphunzitsa nsanja yake ya maonekedwe kuyambira kutulutsidwa kwake koyamba kotseguka, ndicho chifukwa chake kumvetsa zikalata, kumvetsa UI ndi zithunzi za skrini, ndi kuwerenga ma chart zakhala mphamvu zokhazikika za banjali osati zoganiziridwa pambuyo pake.

MiniMax-M1 — chisamalira chosakanikirana cha kuganiza kwautali

MiniMax-M1 (arXiv 2506.13585) unafotokozedwa ndi olemba ake ngati mtundu woyamba wa zolemera zotseguka, wa mlingo waukulu, woganiza wa chisamalira chosakanikirana. Unasunga thupi la MoE la 456B / 45.9B ndipo unaluka mitundu iwiri ya chisamalira: block ya transformer ya softmax-attention yokhazikika pambuyo pa ma block asanu ndi awiri a Lightning Attention. Phindu linayezedwa mu masamu m'malo mwa malo pa mndandanda — MiniMax inanena kuti pa kutalika kwa kupanga kwa ma token 100K, M1 inagwiritsa ntchito pafupifupi 25% ya ma FLOP a DeepSeek R1, pomwe ikuthandiza kuyambira pachiyambi context ya miliyoni ya ma token.

MiniMax-M2 — kutembenukira ku kuyatsa kochepa

MiniMax-M2 unatenga njira ina pa kukula ndipo unapeza chidwi chifukwa cha izo: decoder ya zosanjikiza 62 ndi zigawo zonse 229.9B koma ~9.8B zokha zogwira ntchito pa token iliyonse, pogwiritsa ntchito akatswiri 256 ang'onoang'ono pa chosanjikiza chilichonse cha MoE ndi routing ya sigmoid kupita kwa asanu ndi atatu a iwo, malo a RoPE ndi QK-RMSNorm pa kukhazikika kwa kuphunzitsa. Unali mtundu wa agentic-coding mwachindunji — kusintha kwa mafayilo ambiri, kuzungulira kwa kuyendetsa-ndi-kukonza, kukonza kotsimikiziridwa ndi mayeso — ndipo unayambitsa mawonekedwe a kuganiza kolukidwa, kumene kuganiza kumanyamulidwa pakati pa mizere ya kuyitanitsa zida m'malo mochotsedwa pambuyo pa mzere uliwonse. M2 ndi mtundu umene unakhazikitsa mbiri ya MiniMax yoperekera khalidwe la agentic loyandikira pamwamba pa gawo laling'ono la mtengo pa token, mwa kusunga kuyatsidwa kokhala kochepa kwambiri.

MiniMax-M3 — maziko a endpoint iyi

MiniMax-M3 unatulutsidwa pa June 1, 2026, ndi zolemera pa Hugging Face patapita nthawi yochepa ndi lipoti la ukadaulo pa arXiv 2606.13392. Ndi mtundu wa Mixture-of-Experts wa zigawo 428B ndi pafupifupi 23B zogwira ntchito pa token iliyonse, context yolembedwa ya miliyoni ya ma token, ndi zolowetsa za zithunzi ndi makanema kuyambira pachiyambi — MiniMax imaufotokoza ngati wophunzitsidwa mwa mitundu yosakanikirana "kuyambira sitepe yoyamba", m'malo mwa kuphunzitsidwa pa mawu kenako kusinthidwa kukhala wa multimodal. Pa manambala ake ofalitsidwa a MiniMax umapeza 80.5% pa SWE-bench Verified ndi 59.0% pa SWE-bench Pro. Amenewo ndi ziwerengero zonenedwa ndi wogulitsa pa benchmark ya code, ndipo ndendende ndi mmene muyenera kuwalemekezera; timawatchula chifukwa ndi zonena zofalitsidwa za labotale, osati chifukwa tinawabwereza tokha.

03MiniMax Sparse Attention: chifukwa chake mtundu wa masomphenya ungakhale wotsika mtengo chonchi

Uinjiniya wosangalatsa mu M3 ndi MiniMax Sparse Attention (MSA), ndipo ndicho chifukwa chenicheni chimene mtundu wa multimodal wa kukula kumeneku ungakhalire pa mtengo wathu wa pansi. Zithunzi zimadya ma token. Masamba ochepa a resolution yayikulu angaike ma token masauzande makumi ambiri mu funso musanalembe langizo limodzi. Pansi pa chisamalira chokhazikika, mtengo umenewo umachuluka mowirikiza, ndipo mtundu umene umawerenga masamba makumi awiri sukhala wokwera mtengo kuwirikiza kawiri kuposa umene umawerenga khumi — umakhala woipa kwambiri kuposa kuwirikiza kawiri.

MSA imasintha mawonekedwe a chikhotocho. Ndi njira ya block-sparse yomangidwa pa msana wa Grouped Query Attention, yogawidwa m'nthambi ziwiri:

• Index Branch yopepuka imene imayerekezera ma block a key-value — ziwerengero za max-pooling pa ma token mkati mwa block iliyonse — ndipo imasankha top-k pa malo a funso lililonse ndi pa gulu lililonse la GQA, nthawi zonse kuphatikiza block ya pafupi imene ili ndi funso lokhalo pa kukhazikika kwa kuphunzitsa.
• Main Branch imene imayendetsa chisamalira cha softmax chokhazikika pa ma block amene index inasankha okha.

Kakonzedwe mu pepala la MSA ndi k = ma block 16 a ma token 128, kumene kumapatsa funso lililonse ndalama zokhazikika za ma token 2,048 osamaliridwa. Zokhazikika, kutanthauza kuti sizikula ndi kutalika kwa mzere. Ndicho chinyengo chonse: pamene context ikukula, mtengo wa chisamalira wa mtundu wolimba wa poyerekezera umapitiriza kukwera pomwe mtengo wa main-branch wa MSA umakhala wosasintha, choncho chisamalira chimakula mwa mzere osati mowirikiza.

Ziwerengero zochokera ku lipoti la MiniMax Sparse Attention (arXiv 2606.13392): kuchepetsa kwa kuwirikiza 28.4× kwa ma FLOP a chisamalira pa token iliyonse poyerekeza ndi GQA yokonzedwa mofanana pa context ya 1M, ndi kufulumira kwa nthawi ya wotchi kwa prefill kwa 14.2× / decode kwa 7.6× kuyezedwa pa zida za H800 mu kafukufuku wa pepalali wa zigawo 109B. MiniMax padera ikunena za kufulumira kwa prefill kwa 9× ndi decode kwa 15× kwa M3 poyerekeza ndi M2 pa context ya 1M. Zonse ndi zonenedwa ndi labotale.

Pepalali limachitanso chiyeso chimene chimafunika pa nkhani ya masomphenya: pa MoE ya 109B ya multimodal kuyambira pachiyambi, MSA inasunga machitidwe ofanana ndi a poyerekezera wa chisamalira chonse pa ma benchmark a zithunzi kuphatikiza AI2D, ChartQA ndi MMMU, ndi pa ma benchmark a makanema kuphatikiza EgoSchema, LongVideoBench ndi VideoMME — ndi mitundu yosakanikirana nthawi zina ikupambana pa ntchito za makanema za context yaitali. Kusakanikirana pano si msonkho wa mtundu umene mumalipira chifukwa cha liwiro; pa zolowetsa zazitali zikuoneka kuti ndi kwaulere.

Chifukwa chake muyenera kusamala ngati kasitomala: mtengo wa kutumikira ndiwo umakhazikitsa mtengo wa mndandanda. Maziko amene mtengo wake wa chisamalira umasiya kukula ndi maziko amene mungathe kuwaloza pa zithunzi zana masauzande. Imeneyo ndi njira imene ili kumbuyo kwa "mtundu wa masomphenya wosasidwa wotsika mtengo kwambiri" — si mtengo wotsatsa, ndi kamangidwe.

04Chomwe 3BIT-REAP chikutanthauza, ndi kumene chikuwonongerani

Mtundu wa zigawo mabiliyoni 428 sukwana pena paliponse posavuta pa kulondola konse. Kupanga kwathu kwa MiniMax-M3 kumanyamula chowonjezera cha 3BIT-REAP, chimene chikufotokoza ntchito ziwiri zosiyana zogwiritsidwa ntchito m'dongosolo linalake.

REAP: chotsani akatswiri musanawapanikize

REAP — Router-weighted Expert Activation Pruning — ndi njira ya kupanikiza kwa MoE kamodzi kokha yochokera ku Cerebras ndi University of Calgary, yofalitsidwa ngati arXiv 2510.13999 ndi yovomerezedwa ku ICLR 2026. Mfundo yake ndi yakuti pa ntchito za kupanga, kudula akatswiri kumapambana kuwaphatikiza: kuphatikiza kumawononga ulamuliro wolondola wa router ndipo kumayambitsa cholakwika chimene simungachibwezeretse. Choncho REAP m'malo mwake imayerekezera katswiri aliyense ndi muyeso wa kufunika wophatikiza mitengo ya geti ya router (kuchuluka ndi kulimba kwa kusankhidwa kwake) ndi ma norm a kuyatsidwa kwa katswiri (kuchuluka kwa zomwe zotuluka zake zimathandiza kwenikweni), kenako imachotsa akatswiri ofunika pang'ono kwambiri kwathunthu.

Olemba akunena za kuchotsa mpaka 50% ya akatswiri pa mitundu yaikulu ngati ya zigawo triliyoni imodzi pomwe akusunga kwakukulu mtundu wa poyerekezera — pa Qwen3-480B-Coder-FP8 pa kudula kwa 50%, 97.6% ya luso la kulemba code losakhala la agentic ndi 96.7% pa SWE-Bench ya agentic zinasungidwa. Amenewo ndi manambala a pepalali pa mitundu ya pepalali, osati athu pa MiniMax-M3; timawatchula kuti muweruze njirayo, osati kunena kuti kuyeza kofanana kunachitika pano.

Zolemera za 3-bit, kachiwiri

Zolemera zopulumuka kenako zimasungidwa pa mabiti 3. Dongosolo ndi lodala ndipo ndicho cholinga cha kuphatikizaku: dulani choyamba, panikizani kachiwiri. Palibe nzeru zowononga mabiti anu atatu amtengo wapatali pa kulemba mokhulupirika akatswiri amene mukufuna kuwachotsa. Dulani mchira wosakanikirana, kenako lolani ndalama zotsalira zipite kwathunthu kwa akatswiri amene amayatsidwadi.

Chenjezo loona

3BIT-REAP ndi mlingo wolimba kwambiri umene timatumiza. Timaugwiritsa ntchito pa maziko akulu kwambiri — MiniMax-M3, Kimi-K3, DeepSeek-V4-Pro, GLM-5.2, Nemotron3Ultra — chifukwa ndicho kusiyana pakati pa mitundu imeneyo kupezeka pa mtengo womveka ndi kusapezeka konse. Kumene maziko ndi ang'onoang'ono mokwanira kuti asafunikire, sitiugwiritsa ntchito: mitundu ya W8A16 mu katalogi imayendetsa zolemera za 8-bit ndi zochitika za 16-bit, ndipo W4A16-AUTOROUND-REAP ili pakati, ikugwiritsa ntchito kuzungulira kophunziridwa kwa AutoRound kubweza kulondola kumene quantization ya kuzungulira ku pafupi mosaganizira imapereka. Sankhani mlingo umene umagwirizana ndi kulolera kwanu. Ngati mukutulutsa magawo mu ma invoice zikwi khumi, 3-bit ndi kusinthana kolondola. Ngati mukuchita kuwerenga kumodzi kwa umboni wa mlandu kwa chithunzi chimodzi chofunika kwambiri, gwiritsani ntchito ndalama pa Kimi-K3.

05Chomwe "palibe wosanjikiza wa kukana" chikutanthauza pamene chilowetso ndi chithunzi

Chosasidwa ndi liwu limene limagwiritsidwa ntchito mosamala pang'ono, choncho nachi chonena chenicheni: palibe wosanjikiza wa kukana ndipo palibe kusefa kwa zotulutsa pa endpoint iyi. Palibe chowunikira chimene chimayang'ana chithunzi chanu mtunduwo usanachione. Palibe mtundu wa ndondomeko umene umayang'ana ma token pa njira yotuluka. Chimene mtunduwo umapanga ndicho chimene mumalandira.

Mu mawu, zotsatira zake zimadziwika. Mu masomphenya, ndi zosiyana mtundu, chifukwa khalidwe la kukana pa zolowetsa za zithunzi ndiko kumene ma API a multimodal amalonda amakhala osamala kwambiri ndi osasankha bwino kwambiri. Zolephera zake zooneka ndi izi:

• Kukana konse kwa kuzindikira anthu — kukana kufotokoza anthu mu chithunzi konse, kuphatikiza makhalidwe amene akuoneka bwino ndipo ndi osavulaza konse, chifukwa ndondomeko singasiyanitse kufotokoza ndi kuzindikiritsa.
• Kuthawa chifukwa cha mutu — kukana kulemba kapena kufupikitsa chikalata chifukwa cha chimene chikukamba, ngakhale kulemba ndi ntchito ya makina.
• Manyazi a zachipatala ndi a umboni wa milandu — kukana kufotokoza zithunzi za chipatala, zolemba za kuvulala, kapena zithunzi za ngozi zimene katswiri akufunika kuti zifotokozedwe.
• Zolakwika zabodza za zida za chitetezo — kukana kuwerenga chithunzi cha skrini chifukwa chili ndi terminal, packet capture, kapena debugger.
• Kuwonongeka kwakachetechete — koipa kwambiri: osati kukana koma yankho losadziwika bwino, limene limadutsa kutsimikizira kwanu kwa schema ndipo limawononga deta yanu popanda kukweza cholakwika konse.

Vuto la ntchito ndi zonse zisanu si loti sizolungama. Ndi loti sizingayezedwe. Yankho lolakwika ndi mfundo ya deta; mungaipatse chiwerengero, kuipeza mu kuyeza, ndi kuikonza. Kukana ndi bowo. Pa mlingo waukulu, mlingo wa kukana wa 3% wogawidwa mosakhala mwangozi pa zolowetsa zanu — wounjikika, mwa kamangidwe, ndendende pa milandu yolimba ndi yosazolowereka — ndi kukondera kokhazikika mu zotsatira zanu kumene kuyeretsa kulikonse kwapambuyo sikungakubwezeretse. Kuchotsa wosanjikiza wa kukana kumachotsa gwero la deta yosowa yakachetechete, yogwirizana.

Shannon imayendetsedwa ndi Shannon Lab LLC, New Mexico, USA, pansi pa Ndondomeko yathu ya Kagwiritsidwe Kaudindo. Chosasidwa chikutanthauza kuti mtunduwo suweruza za zolowetsa zanu; sichikutanthauza kuti akaunti ilibe mfundo.

06Malongosoledwe onse

Gawo	MiniMax-M3 · 3BIT-REAP
Id ya mtundu	MiniMax-M3-3BIT-REAP
Maziko	MiniMax-M3 (MiniMax AI)
Zenera la context	ma token 262,144
Kutuluka kwambiri	ma token 16,384
Zolowetsa za zithunzi	Inde
Zolowetsa za makanema	Sizinaperekedwe
Streaming	Inde
Kuyitanitsa ntchito	Inde
Kuganiza	Inde (reasoning_effort)
Zotuluka zokonzedwa	json_schema + response_format
Wosanjikiza wa kukana	Palibe
Ma endpoint	/v1/chat/completions · /v1/messages · /v1/responses
Yotumikiridwa pa	Gulu la Shannon la ma GPU lokha

Mizere iwiri mu tebulo limenelo ikufuna kutsindikizidwa. Zolowetsa za makanema sizinaperekedwe. Maziko a M3 a kumtunda anaphunzitsidwa pa makanema ndipo MiniMax imalemba za zolowetsa za makanema; endpoint yathu imalandira zithunzi. Mangani pa zithunzi. Chachiwiri, zenera la ma token 262,144 ndi lathu, osati chiwerengero cha kumtunda cha miliyoni imodzi — konzani ntchito za zikalata zazitali ndi za zithunzi zambiri motsutsana ndi 262,144 ndipo simudzadabwa konse.

07Kuyitana API yosasidwa ya MiniMax

Zilankhulo zitatu, mtundu womwewo kumbuyo kwa zonse, streaming pa zonse zitatu. Gwiritsani ntchito chimene kasitomala wanu wa panopa amalankhula kale; palibe kusiyana kwa luso pakati pawo. Lozani base URL yanu ku Shannon ndipo sinthani id ya mtundu.

Kumvetsa zithunzi kudzera mu chinenero cha 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
  }'

Ma URL akutali a https:// ndi ma data URL a base64 onse amagwira ntchito mu gawo la image_url. Matirani zigawo za zithunzi zingapo mu uthenga umodzi pamene mukufuna kuyerekezera masamba osiyanasiyana — muli ndi ma token 262,144 owonongera ndipo, pa $0.50 pa miliyoni, kuwawononga ndi kotsika mtengo.

Kutulutsa deta kwa mtundu wodziwika ndi JSON schema

Kumene mtunduwu umapeza malo ake mu njira ya deta ndi kuphatikiza kwa masomphenya + schema yolimba + mtengo wa pansi. Si mtundu uliwonse mu katalogi yathu uli ndi zonse zitatu: Laguna-S-2.1 ndi inkling sizithandiza json_schema. MiniMax-M3 imathandiza, kutanthauza kuti mungapeze chinthu chotsimikiziridwa kuchokera ku chithunzi popanda wosanjikiza wa kusanthula:

{
  "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
      }
    }
  }
}

Khama la kuganiza

Mtunduwo umapereka reasoning_effort. Pa kuika m'magulu zithunzi za bulki, likhale lotsika — mukufunsa funso la kuzindikira, ndipo ma token a kuganiza ndi ma token otuluka amene mumalipira. Likwezeni pa mayitanidwe kumene chithunzi ndi umboni mu mtsutso wautali: kuwerenga chithunzithunzi motsutsana ndi malongosoledwe, kugwirizanitsa chart ndi tebulo, kusankha ngati zithunzi ziwiri za skrini zikusonyeza cholakwika chomwecho.

Mndandanda wonse wa ma parameter, tanthauzo la zolakwika, ndi mawonekedwe a streaming a zilankhulo zonse zitatu zili mu zolemba za API ya Shannon.

08Kumene MiniMax-M3 ndi chisankho choyenera — ndi kumene siili

Ifikireni pamene

• Ntchito ya zithunzi za kuchuluka kwakukulu. Kuika zilembo pa dataset, kafukufuku wa kuyang'anira, kulemeretsa katalogi, kusanthula zithunzi za skrini, OCR-ndi-mawonekedwe a bulki pa zikalata zosindikizidwa. Mtengo wa pansi ndi mfundo yonse.
• Zithunzi kupita ku mbiri za mtundu wodziwika. Masomphenya kuphatikiza json_schema yolimba kuphatikiza ma token otsika mtengo kwambiri mu katalogi ndi kuphatikiza kumene mitundu iwiri yokha pano imapereka.
• Context yaitali ya multimodal. Masamba ambiri, kapena ma frame ambiri, mu pempho limodzi — kamangidwe kake ka pansi kanamangidwira ndendende mawonekedwe amenewa a chilowetso.
• Kafukufuku kumene kukana kumawononga chitsanzo. Kafukufuku aliyense amene kutsimikizika kwake kumadalira kuti mtunduwo uyankhe chinthu chilichonse, kuphatikiza zosasangalatsa.
• Kuzungulira kwa agent kumene kumawerenga skrini. Mzere wa M ndi wa agentic mwachindunji, ndipo kuyitanitsa ntchito kulipo.

Fikirani china pamene

• Mukufuna malire apamwamba kwambiri a masomphenya ndipo ndalama zilipo. Kimi-K3 · 3BIT-REAP ndi kuwirikiza 7.7× kwa mtengo wa kulowa pa chifukwa; pa kuwerenga kumodzi kofunika kwambiri, imeneyo ndi ndalama yoyenera kuwononga.
• Mukufuna zolemera za kukhulupirika kwapamwamba kwambiri. 3-bit ndi mlingo wathu wolimba kwambiri. Mitundu ya W8A16 imasinthanitsa mtengo ndi kulondola, ndipo MiMo-V2.5 · W8A16 imakhoza masomphenya pa $0.50 / $2.00 omwewo ngati kukhulupirika pa kulemera ndicho chimene mukukonza.
• Mukufuna makanema. Sitipereka zolowetsa za makanema pa endpoint iyi.
• Ntchito ndi ya mawu okha pa mlingo wapamwamba. DeepSeek-V4-Pro-0813 ilipo pa izo; mukanakhala mukulipira nsanja ya masomphenya imene simuigwiritsa ntchito.

Ziwerengero za kamangidwe, benchmark, ndi kufulumira za banja la MiniMax ndi zonenedwa ndi labotale ya MiniMax AI; ziwerengero za REAP zanenedwa ndi olemba a REAP pa mitundu yawo yokha ndipo si muyeso wa kupanga uku. Mitengo ya Shannon, zenera la context, ndi zizindikiro za luso ndi zathu ndipo zawerengedwa kuchokera ku katalogi yathu yeniyeni ya mitundu.

09Mafunso ofunsidwa kawirikawiri

---

Magwero: Kaadi ya mtundu ya 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. Ziwerengero za benchmark ndi za kufulumira za MiniMax ndi zonenedwa ndi labotale ndipo ndi zosakhazikika zikudikira kutsimikizira kodziyimira pawokha. Mitengo ya Shannon ndi zizindikiro za luso ndi zapano pofika Seputembala 5, 2026.