API na MiniMax marar tantancewa: gani a farashi mafi ƙasa

Kusan kowace API mai yawan hanyoyi da ake hidimarta tana sanya farashin gani a matsayin abu mai tsada. Kana biyan kuɗin samfurin rubutu don rubutu, sannan ka biya ƙari — wani lokaci ninki da dama — a daidai lokacin da hoto ya shiga cikin umarnin. Bugu da ƙari, shigar hoto shi ne inda ɗabi'ar ƙi ta fi tsanani: samfurin da zai tattauna wani batu cikin rubutu cikin farin ciki zai ƙi kwatanta abu ɗaya a cikin hoto. Sakamakon shi ne kayan da yake da tsada a kowane kira kuma marar aminci a kowane kira, wanda shi ne mafi muni a fannoni biyu idan kana sarrafa yawa. MiniMax-M3 a Shannon ita ce amsarmu ga hakan: $0.50 / $2.00 guda da kake biya don samfurin rubutu mafi rahusa, an yi amfani da shi a kan tushen da aka horar da shi mai yawan hanyoyi tun mataki na farko, ba tare da shimfidar ƙi a gabansa ba.

01Menene ma'anar "samfurin gani marar tantancewa mafi rahusa" a nan

Katalogin Shannon samfura goma sha biyu ne. Biyar daga cikinsu suna karɓar hotuna. Farashinsu, kowane tokens miliyan:

Samfuri mai iya gani	Shigarwa	Fitarwa	Fitarwa mai tsari
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 kaɗai
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	json_schema + response_format

MiniMax-M3 yana zaune a mafi ƙasa — $0.50 / $2.00 — daidai da MiMo-V2.5 kuma babu wani a ƙasansa, saboda $0.50 / $2.00 shi ne ƙasan jerin farashinmu. Idan aka kwatanta da saman fagen gani, ya fi rahusa sau 7.7 a shigarwa da sau 9.6 a fitarwa fiye da Kimi-K3.

Ka sanya hakan a cikin aiki. A ce kana tantance hotuna 1,000 — hotunan allo, shafuka da aka duba, hotunan kaya — matsakaicin tokens 1,200 na hoto tare da umarni suna shiga da tokens 300 na amsa suna fita:

Farashin jerin Shannon, Satumba 2026. Adadin tokens misali ne kawai; rarraba hotonka zuwa tokens ya dogara ga ƙudurin hoto da yawan hotunan da ka haɗa a kowane kira.

Wasu maki biyu na tunani da suke da amfani ka riƙe a kai. Cika duka mahalli na tokens 262,144 sau ɗaya yana kashe kusan cent 13. Amsa mai matsakaicin tsawo na tokens 16,384 tana kashe kusan cent 3.3. Waɗannan su ne ainihin iyakokin buƙata guda mafi tsada, kuma ƙananan ne har ƙuntatawar da take da muhimmanci a kan hanyar aikin gani ta daina zama kuɗi ta zama yawan aiki.

02Zuriyar MiniMax: dakin bincike da yake inganta tokens, ba kanun labarai ba

MiniMax AI dakin bincike ne na Shanghai wanda duka tarihin nauyinsa buɗaɗɗe yake game da tambaya guda: ta yaya za ka sa kulawa ta yi rahusa har dogon mahalli da yawan hanyoyi su daina zama alatu? Kowane saki a wannan layin amsa ce ga wannan tambayar, kuma amsoshin suna gina a kan juna.

MiniMax-01 — Lightning Attention, da gani tun daga farko

Jerin MiniMax-01 (an wallafa a matsayin arXiv 2501.08313) ya gabatar da Lightning Attention, aiwatarwa mai lura da I/O na wani nau'in kulawa madaidaiciya, a cikin samfurin Mixture-of-Experts haɗe mai jimlar sigogi biliyan 456 da biliyan 45.9 masu kunnuwa a kowane token. Jerin ya fito biyu-biyu: MiniMax-Text-01 da MiniMax-VL-01, samfurin gani. VL-01 ya yi amfani da tarin ViT-MLP-LLM da aka saba da shi yanzu, amma da Vision Transformer mai sigogi miliyan 303 da aka horar tun daga sifili a kan nau'i-nau'i miliyan 694 na hoto-da-bayani maimakon mai ɓoyewa da aka aro, sannan aka daidaita shi kuma aka tace shi tare a kan hanyar da take sarrafa tokens biliyan 512.

Wannan cikakken bayanin yana da muhimmanci ga labarin da kake karantawa. Gani a wannan zuriyar ba wani mai daidaitawa ne da aka manne a kan samfurin rubutu daga baya don a cika akwati ba. MiniMax yana horar da hasumiyar ganinsa tun sakinsa buɗaɗɗe na farko, wanda shi ne ya sa fahimtar takardu, fahimtar UI da hotunan allo, da karanta zane suka kasance ƙarfin dangin koyaushe maimakon abin da aka yi tunani a kai daga baya.

MiniMax-M1 — kulawa haɗe don tunani a tsawo

MiniMax-M1 (arXiv 2506.13585) marubutansa sun kwatanta shi da cewa samfurin tunani na farko mai nauyi buɗaɗɗe, mai girma, mai kulawa haɗe. Ya riƙe jikin MoE na 456B / 45.9B kuma ya haɗa nau'ikan kulawa biyun: toshen transformer na kulawar softmax na yau da kullum bayan kowane toshen Lightning Attention bakwai. An auna ribar da lissafi maimakon matsayi a jerin — MiniMax ya bayar da rahoton cewa a tsawon samarwa na tokens 100K, M1 ya cinye kusan kashi 25% na FLOPs na DeepSeek R1, alhali yana goyon bayan mahalli na tokens miliyan ɗaya ta asali.

MiniMax-M2 — juyi zuwa ƙarancin kunnuwa

MiniMax-M2 ya bi wata hanya daban a girma kuma ya sami kulawa saboda hakan: mai ɓoyewa mai shimfidu 62 mai jimlar sigogi biliyan 229.9 amma kusan biliyan 9.8 kaɗai masu kunnuwa a kowane token, yana amfani da masana 256 masu daki-daki a kowace shimfidar MoE tare da turawa ta sigmoid zuwa takwas daga cikinsu, matsayin RoPE da QK-RMSNorm don kwanciyar hankalin horo. A fili samfurin rubuta lamba na wakilai ne — gyaran fayiloli da dama, madaukan gudu-da-gyara, gyara da gwaji ya tabbatar — kuma ya gabatar da tsarin tunani mai haɗewa, inda ake ɗaukar tunani tsakanin zagayen kiran kayan aiki maimakon a jefar da shi bayan kowane juyi. M2 shi ne samfurin da ya kafa sunan MiniMax na isar da ɗabi'ar wakilai kusa da gaba a kaso ɗaya na kuɗin kowane token, kawai ta hanyar riƙe kunnuwa ƙanana.

MiniMax-M3 — tushen da ke ƙarƙashin wannan ƙarshen hanyar

An saki MiniMax-M3 a 1 ga Yuni, 2026, tare da nauyi a Hugging Face jim kaɗan bayan haka da rahoton fasaha a arXiv 2606.13392. Samfurin Mixture-of-Experts ne mai sigogi biliyan 428 da kusan biliyan 23 masu kunnuwa a kowane token, mahalli na tokens miliyan ɗaya da aka rubuta, da shigar hoto da bidiyo na asali — MiniMax ya kwatanta shi da cewa an horar da shi da hanyoyi gauraye "tun daga mataki na farko kwata-kwata," maimakon a horar da shi da rubutu sannan a daidaita shi da yawan hanyoyi daga baya. A kan lambobin da MiniMax ya wallafa da kansa yana samun 80.5% a SWE-bench Verified da 59.0% a SWE-bench Pro. Waɗannan lambobin mai kera ne ya bayar da rahotonsu a kan ma'aunin rubuta lamba, wanda shi ne yadda ya kamata ka auna su daidai; muna ambatonsu saboda iƙirarin da dakin binciken ya wallafa ne, ba don mun sake samar da su da kanmu ba.

03MiniMax Sparse Attention: dalilin da ya sa samfurin gani zai iya zama mai wannan rahusar

Aikin injiniya mai ban sha'awa a M3 shi ne MiniMax Sparse Attention (MSA), kuma shi ne dalili kai tsaye da ya sa samfuri mai yawan hanyoyi na wannan girman zai iya zama a farashinmu mafi ƙasa. Hotuna suna cin tokens. Ƴan shafuka masu ƙuduri mai girma za su iya sanya dubun dubatar tokens cikin umarni kafin ka rubuta umarni guda. A ƙarƙashin kulawa ta yau da kullum, wannan kuɗin yana ninkuwa sau biyu, kuma samfurin da ya karanta shafuka ashirin ba ninki biyu ne na wanda ya karanta goma ba — ya fi ninki biyu muni sosai.

MSA yana canja siffar wannan lanƙwasan. Hanya ce mai raguwar toshe da aka gina a kan Grouped Query Attention, an raba ta zuwa rassa biyu:

• Reshen Fihirisa mai sauƙi wanda yake ba wa toshen maɓalli-da-daraja maki — yana tara mafi girman maki a cikin tokens na kowane toshe — sannan ya zaɓi na sama-k a kowane matsayin tambaya da kowace ƙungiyar GQA, koyaushe yana haɗa toshen gida da ke ɗauke da tambayar kanta don kwanciyar hankalin horo.
• Babban Reshe wanda yake gudanar da kulawar softmax ta yau da kullum kawai a kan toshen da fihirisar ta zaɓa.

Tsarin da yake cikin takardar MSA shi ne k = toshe 16 na tokens 128, wanda yake ba kowace tambaya kasafi tsayayye na tokens 2,048 da ake kulawa da su. Tsayayye, ma'ana ba ya girma da tsawon jerin. Wannan shi ne duka dabarar: yayin da mahalli ke girma, kuɗin kulawa na ma'auni mai yawa yana ci gaba da hawa alhali kuɗin babban reshen MSA ya kasance a tsaye, don haka kulawa tana girma madaidaiciya maimakon ninki biyu.

Lambobi daga rahoton MiniMax Sparse Attention (arXiv 2606.13392): raguwar sau 28.4 a FLOPs na kulawa a kowane token idan aka kwatanta da GQA mai tsari daidai a mahalli na 1M, da saurin cikawa sau 14.2 / ɓoye-ɓoye sau 7.6 na lokacin agogo da aka auna a kan na'urar H800 a cikin binciken takardar na sigogi biliyan 109. MiniMax ya bayar da rahoto daban na saurin cikawa sau 9 da ɓoye-ɓoye sau 15 ga M3 idan aka kwatanta da M2 a mahalli na 1M. Duka dakin bincike ne ya bayar da rahotonsu.

Takardar ta kuma yi dubawar da take da muhimmanci ga labarin gani: a kan MoE mai sigogi biliyan 109 mai yawan hanyoyi na asali, MSA ya riƙe aiki kwatankwacin ma'auni na cikakkiyar kulawa a kan ma'aunan hoto haɗe da AI2D, ChartQA da MMMU, da kuma a kan ma'aunan bidiyo haɗe da EgoSchema, LongVideoBench da VideoMME — inda nau'ikan masu raguwa wani lokaci suke gyara ayyukan bidiyo masu dogon mahalli. Raguwa a nan ba haraji ne na inganci da kake biya don sauri ba; a kan shigarwa masu tsawo yana kama da kusan kyauta.

Dalilin da ya sa hakan ya kamata ya damu ka a matsayin abokin ciniki: kuɗin hidima shi ne yake sanya farashin jeri. Tushen da kuɗin kulawarsa ya daina girma shi ne tushen da za ka iya nunawa ga hotuna dubu ɗari. Wannan ita ce hanyar da ke ƙarƙashin "samfurin gani marar tantancewa mafi rahusa" — ba farashin talla ba ne, gine-gine ne.

04Menene ma'anar 3BIT-REAP, kuma a ina yake kashe maka kuɗi

Samfuri mai sigogi biliyan 428 ba ya shiga a wani wuri mai dacewa a cikakken daidaito. Ginin MiniMax-M3 namu yana ɗauke da ƙarin 3BIT-REAP, wanda yake kwatanta ayyuka biyu daban da aka yi da tsari na musamman.

REAP: goge masana kafin ka matsa su

REAP — Router-weighted Expert Activation Pruning — hanyar matsi na MoE ce ta harbi guda daga Cerebras da Jami'ar Calgary, an wallafa ta a matsayin arXiv 2510.13999 kuma an karɓe ta a ICLR 2026. Hujjarta ita ce ga ayyuka na samarwa, datsa masana ya fi haɗa su: haɗawa yana lalata ikon router mai daki-daki kuma yana gabatar da kuskuren da ba za ka taɓa dawo da shi ba. Don haka REAP maimakon haka yana ba kowane masani maki da ma'aunin muhimmanci da yake haɗa darajojin ƙofar router (sau nawa kuma da ƙarfin nawa router yake zaɓarsa) da ma'aunin kunnuwan masani (nawa fitarwarsa take bayarwa da gaske), sannan ya goge waɗanda suka fi ƙaranci kai tsaye.

Marubutan sun bayar da rahoton cire har zuwa kashi 50% na masana daga samfura masu girman sigogi tiriliyan ɗaya alhali sun kiyaye ingancin asali gaba ɗaya — a kan Qwen3-480B-Coder-FP8 a datsawar 50%, an riƙe 97.6% na iyawar rubuta lamba mara ta wakilai na asali da 96.7% a SWE-Bench na wakilai. Waɗannan lambobin takardar ne a kan samfuran takardar, ba namu a kan MiniMax-M3 ba; muna ambatonsu don ka iya yanke hukunci a kan hanyar, ba don mu nuna cewa akwai auni makamancin haka a nan ba.

Nauyi na 3-bit, na biyu

Sannan ana adana nauyin da ya tsira a 3 bits. Tsarin da gangan ne kuma shi ne ma'anar haɗin: datse da farko, matsa na biyu. Babu ma'ana a kashe bits ɗinka uku masu daraja a kowane nauyi kana ɓoye masana da za ka goge nan gaba da aminci. Ka datse wutsiyar da ba ta cika ba, sannan ka bar sauran kasafin ya tafi gaba ɗaya ga masanan da suke aiki da gaske.

Gargaɗi na gaskiya

3BIT-REAP shi ne matakin da ya fi tsanani da muke bayarwa. Muna amfani da shi a kan manyan tushe — MiniMax-M3, Kimi-K3, DeepSeek-V4-Pro, GLM-5.2, Nemotron3Ultra — saboda shi ne bambanci tsakanin waɗannan samfuran su kasance a farashi mai ma'ana ko kuma kada su kasance kwata-kwata. Inda tushe ya yi ƙanƙanta har ba ya buƙatarsa, ba ma amfani da shi: samfuran W8A16 a katalogin suna gudanar da nauyi na 8-bit da kunnuwa na 16-bit, kuma W4A16-AUTOROUND-REAP yana zaune tsakaninsu, yana amfani da zagayen da AutoRound ya koya don dawo da daidaiton da matsi na zagaye zuwa mafi kusa mai sauƙi yake bayarwa. Ka zaɓi matakin da ya dace da haƙurinka. Idan kana cire filayen daga takardun kuɗi dubu goma, 3-bit shi ne ciniki daidai. Idan kana yin karatun shaida guda mai matuƙar muhimmanci na hoto ɗaya, ka kashe kuɗin a kan Kimi-K3.

05Menene ma'anar "babu shimfidar ƙi" idan shigarwar hoto ce

Marar tantancewa kalma ce da ake amfani da ita a sassauce, don haka ga takamaiman iƙirarin: babu shimfidar ƙi kuma babu matatar fitarwa a wannan ƙarshen hanyar. Babu wani mai rarrabewa da yake duba hotonka kafin samfurin ya gan shi. Babu samfurin manufa da yake duba tokens a hanyar fita. Abin da samfurin ya samar shi ne abin da ka karɓa.

A cikin rubutu, sakamakon hakan sananne ne. A cikin gani, ya bambanta a nau'i, saboda ɗabi'ar ƙi a kan shigar hoto ita ce inda API na kasuwanci masu yawan hanyoyi suka fi taka tsantsan kuma suka fi rashin tsari. Kurakuran da suka fi bayyana suna kama da haka:

• Ƙin ganewa gaba ɗaya — ƙin kwatanta mutane a cikin hoto kwata-kwata, haɗe da halayen da suke bayyane a sarari kuma marasa illa gaba ɗaya, saboda manufa ba za ta iya bambanta kwatantawa da ganewa ba.
• Gudu daga batun — ƙin rubutawa ko taƙaita takarda saboda abin da takardar take magana a kai, ko da yake rubutawa aiki ne na inji.
• Jin kunya na likita da na shaida — ƙin kwatanta hotunan asibiti, takardun rauni, ko hotunan haɗari da ƙwararre yake buƙatar a kwatanta masa.
• Kuskuren tabbatarwa na kayan aikin tsaro — ƙin karanta hoton allo saboda yana ɗauke da tasha, kama fakiti, ko mai gyara lamba.
• Raguwa a shiru — mafi muni: ba ƙi ba ne amma amsa mai kaucewa, wadda take wucewa ta tabbatar da schema ɗinka kuma tana gurɓata bayananka ba tare da ta taɓa ɗaga kuskure ba.

Matsalar aiki da duka biyar ba ita ce rashin adalcinsu ba. Ita ce ba a iya auna su ba. Amsa mara kyau bayanai ce; za ka iya ba ta maki, ka kama ta a lokacin auna, ka gyara ta. Ƙi rami ne. A girman tsari, adadin ƙi na 3% da aka rarraba ba bisa tsari ba a kan shigarwarka — an tattara shi, ta gine-gine, a kan lamurra masu wuya da ba a saba ba daidai — son zuciya ne na tsari a cikin sakamakonka wanda babu tsaftacewa nawa daga baya da za ta dawo da shi. Cire shimfidar ƙi yana cire tushen bayanan da suka ɓace a shiru kuma cikin haɗin kai.

Shannon Lab LLC, New Mexico, Amurka, ce take gudanar da Shannon, ƙarƙashin Manufarmu ta Amfani da Alhaki. Marar tantancewa yana nufin samfurin ba ya yin sharhi a kan shigarwarka; ba ya nufin asusun ba shi da sharuɗɗa.

06Cikakken ƙayyadaddun bayanai

Filin	MiniMax-M3 · 3BIT-REAP
Alamar samfuri	MiniMax-M3-3BIT-REAP
Tushe	MiniMax-M3 (MiniMax AI)
Taga mahalli	tokens 262,144
Matsakaicin fitarwa	tokens 16,384
Shigar hoto	Eh
Shigar bidiyo	Ba a buɗe ba
Streaming	Eh
Kiran kayan aiki	Eh
Tunani	Eh (reasoning_effort)
Fitarwa mai tsari	json_schema + response_format
Shimfidar ƙi	Babu
Filayen API	/v1/chat/completions · /v1/messages · /v1/responses
Ana hidimarsa a kan	Gungun GPU na Shannon

Layuka biyu a wannan teburin sun cancanci a jaddada su. Ba a buɗe shigar bidiyo ba. An horar da tushen M3 na sama a kan bidiyo kuma MiniMax ya rubuta shigar bidiyo; ƙarshen hanyarmu tana karɓar hotuna. Ka gina a kan hotuna. Na biyu, taga na tokens 262,144 namu ne, ba adadin miliyan ɗaya na sama ba — ka tsara ayyukan takardu masu tsawo da hotuna da yawa a kan 262,144 kuma ba za a taɓa ba ka mamaki ba.

07Kiran API na MiniMax marar tantancewa

Yaruka uku, samfuri guda a bayan dukansu, streaming a kan ukun. Ka yi amfani da wanda abokin cinikinka na yanzu ya riga ya sani; babu bambancin iyawa a tsakaninsu. Ka nuna URL ɗinka na asali zuwa Shannon ka canja alamar samfurin.

Fahimtar hoto ta yaren 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
  }'

URL na nesa na https:// da URL na bayanan base64 duka suna aiki a ɓangaren image_url. Ka haɗa ɓangarorin hoto da dama a saƙo ɗaya idan kana buƙatar kwatanci tsakanin shafuka — kana da tokens 262,144 da za ka kashe kuma, a $0.50 kowane miliyan, kashe su abu ne mai rahusa.

Cirar bayanai mai nau'i da JSON schema

Inda wannan samfurin yake cin matsayinsa a kan hanyar aiki shi ne haɗin gani + schema mai tsauri + farashi mafi ƙasa. Ba kowane samfuri a katalogin mu ne yake da ukun ba: Laguna-S-2.1 da inkling ba sa goyon bayan json_schema. MiniMax-M3 yana goyon bayansa, wanda ke nufin za ka iya samun abu da aka tabbatar daga hoto ba tare da shimfidar tantancewa ba:

{
  "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
      }
    }
  }
}

Ƙoƙarin tunani

Samfurin yana buɗe reasoning_effort. A kan rarraba hotuna da yawa, ka riƙe shi ƙasa — kana yin tambaya ta gani, kuma tokens na tunani tokens ne na fitarwa da kake biyan kuɗinsu. Ka ɗaga shi ga kiran da hoton yake shaida a cikin hujja mai tsawo: karanta zane a kan wani ƙayyadaddun bayani, daidaita zane da tebur, yanke hukuncin ko hotunan allo biyu suna nuna kuskure guda.

Cikakken bayanin sigogi, ma'anonin kuskure, da tsarin streaming ga yarukan ukun suna cikin takardun API na Shannon.

08Inda MiniMax-M3 shi ne zaɓi daidai — da inda ba haka ba

Ka kai gare shi lokacin da

• Aikin hoto mai yawa. Yiwa bayanai lakabi, binciken tantancewa, wadata katalogi, tantance hotunan allo, OCR-da-tsari mai yawa a kan takardun da aka duba. Farashi mafi ƙasa shi ne duka hujjar.
• Hotuna zuwa bayanai masu nau'i. Gani tare da json_schema mai tsauri tare da tokens mafi rahusa a katalogin haɗi ne da samfura biyu kaɗai a nan suke bayarwa.
• Dogon mahalli mai yawan hanyoyi. Shafuka da dama, ko firam da dama, a buƙata guda — an gina gine-ginen da ke ƙasa daidai don wannan siffar shigarwa.
• Bincike inda ƙi yake gurɓata samfurin. Duk wani bincike da ingancinsa ya dogara ga samfurin ya amsa kowane abu, haɗe da masu ba da damuwa.
• Madaukan wakilai da suke karanta allo. Zuriyar M-series a fili ta wakilai ce, kuma ana samun kiran kayan aiki.

Ka kai ga wani abu daban lokacin da

• Kana buƙatar mafi girman rufin gani kuma kuɗin yana nan. Kimi-K3 · 3BIT-REAP ya ninka farashin shigarwa sau 7.7 saboda dalili; a kan karatu guda mai matuƙar muhimmanci, wannan shi ne kuɗin da ya dace a kashe.
• Kana buƙatar nauyi mai mafi girman aminci. 3-bit shi ne matakinmu mafi tsanani. Samfuran W8A16 suna sayar da farashi don daidaito, kuma MiMo-V2.5 · W8A16 yana iya gani a $0.50 / $2.00 guda idan aminci a kowane nauyi shi ne abin da kake inganta.
• Kana buƙatar bidiyo. Ba ma buɗe shigar bidiyo a wannan ƙarshen hanyar ba.
• Aikin rubutu ne kawai a gaba. DeepSeek-V4-Pro-0813 yana nan don haka; za ka kasance kana biyan kuɗin hasumiyar gani da ba za ka taɓa amfani da ita ba.

Lambobin gine-gine, ma'auni da sauri ga dangin MiniMax dakin bincike na MiniMax AI ne ya bayar da rahotonsu; marubutan REAP ne suka bayar da rahoton lambobin REAP a kan samfuransu kuma ba auni ne na wannan ginin ba. Farashin Shannon, taga mahalli, da alamun iyawa namu ne kuma ana karanta su daga katalogin samfuranmu na kai tsaye.

09Tambayoyin da ake yawan yi

---

Majiyoyi: Katin samfuri na 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. Dakin bincike ne ya bayar da rahoton lambobin ma'auni da sauri na MiniMax kuma na wucin gadi ne har sai an tabbatar da su da kai. Farashin Shannon da alamun iyawa na yanzu ne tun 5 ga Satumba, 2026.