API MiniMax Kore-ārai: te tirohanga i te papa utu

Ka utua e te nuinga o ngā API reo-maha e whakahaerea ana te tirohanga hei mea utu-teitei. Ka utua koe i te reiti tauira-kuputuhi mō te kuputuhi, kātahi ka utu koe i te nui ake — i ētahi wā he maha whakarea — i te wā ka uru tētahi whakaahua ki te tohutohu. I tua atu i tērā, ko ngā tāuru whakaahua te wāhi kaha rawa te whanonga whakakāhore: ko te tauira e harikoa ana ki te kōrero mō tētahi kaupapa i te kuputuhi ka whakakāhore ki te whakaahua i taua mea ōrite i tētahi whakaahua. Ko te hua he hua utu-nui ia karanga, kore-whakawhirinaki hoki ia karanga, koirā te kino rawa o ngā ao e rua ina tukatuka koe i te ranea. Ko MiniMax-M3 i runga i a Shannon tā mātou whakautu ki tērā: ko te $0.50 / $2.00 ōrite e utua ana e koe mō tā mātou tauira kuputuhi utu-iti rawa, e whakamahia ana ki tētahi tūāpapa i whakangungua reo-maha mai i te hipanga tuatahi, me te kore o tētahi paparanga whakakāhore e noho ana i mua i a ia.

01He aha tonu te tikanga o "te tauira tirohanga kore-ārai utu-iti rawa" i konei

He tekau mā rua ngā tauira o te rārangi Shannon. E rima o rātou ka whakaae ki ngā whakaahua. Anei ō rātou utu, ia miriona tohu:

Tauira whai-tirohanga	Tāuru	Whakaputa	Putanga hanganga
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 anake
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	json_schema + response_format

E noho ana a MiniMax-M3 i te papa — $0.50 / $2.00 — e taurite ana ki a MiMo-V2.5, kāore he mea i raro iho, nā te mea ko te $0.50 / $2.00 te papa o tā mātou rārangi utu. Hei whakataurite ki te pito teitei o te awhe tirohanga, he 7.7× te iti ake i te tāuru, 9.6× te iti ake i te whakaputa i a Kimi-K3.

Whakaurua tērā ki tētahi kawenga mahi. Me whakapae kei te tirotiro tuatahi koe i te 1,000 whakaahua — hopukanga mata, whārangi kua matawaihia, whakaahua hua — he tata 1,200 tohu o te whakaahua me te tohutohu e uru ana, he 300 tohu o te whakautu e puta mai ana, hei toharite:

He utu rārangi nā Shannon, Hepetema 2026. He whakaahuatanga noa ngā tatauranga tohu; ka whakawhirinaki tō tokenization whakaahua ki te iarere me te maha o ngā whakaahua ka whakapiritia e koe ia karanga.

E rua ake ngā tohu tohutoro e whaihua ana kia mau ki tō hinengaro. He tata 13 heneti te utu o te whakakī i te horopaki 262,144-tohu katoa kotahi tau. He tata 3.3 heneti te utu o tētahi whakautu 16,384-tohu roa-nui-rawa. Koirā ngā tino rohe o tētahi tono kino-rawa kotahi, ā, he iti rawa kia mutu ai te rohe whakamīharo o tētahi ara tirohanga hei moni, ka tīmata hei tere whakahaere.

02Te whakapapa MiniMax: he whare rangahau e whakapai ana i ngā tohu, kaua i ngā pane pāpaho

He whare rangahau a Shanghai a MiniMax AI, ā, ko tāna rekoata taumaha-tuwhera katoa e pā ana ki te pātai kotahi: me pēhea e whakaiti utu ai te aro kia mutu ai te horopaki-roa me te reo-maha te noho hei mea whai-rawa noa? He whakautu ki taua pātai ia putanga o te raupapa, ā, ka hanga ngā whakautu i runga i tētahi i tētahi.

MiniMax-01 — Lightning Attention, me te tirohanga mai i te tīmatanga

I whakaurua e te raupapa MiniMax-01 (i whakaputaina hei arXiv 2501.08313) te Lightning Attention, he whakatinanatanga mōhio-ki-te-I/O o tētahi momo aro rārangi, kei roto i tētahi tauira Whakaranu-o-Ngā-Mātanga whakaranu 456B tawhā katoa me te 45.9B hohe ia tohu. I tukuna te raupapa hei takirua: MiniMax-Text-01 me MiniMax-VL-01, te tauira tirohanga. I whakamahia e VL-01 te tū ViT-MLP-LLM mōhiotia nei, engari me tētahi Vision Transformer 303-miriona-tawhā i whakangungua mai i te tīmatanga ki te 694 miriona takirua whakaahua-tapanga, kaua he pūwhakawaehere nama, kātahi ka whakahāngaitia, ka whakatikatikahia tahi puta noa i tētahi ara wāhanga-maha e tukatuka ana i te 512 piriona tohu.

He nui taua taipitopito mō te tuhinga e pānui nei koe. Ehara te tirohanga o tēnei whakapapa i tētahi taputapu whakaurunga tōmuri i herea ki tētahi tauira kuputuhi hei whakaū pouaka noa. Kua whakangungua e MiniMax tāna ake pou tirohanga mai i tāna putanga tuwhera tuatahi, koinā te take kua tū tonu te māramatanga tuhinga, te mātau UI me te hopukanga mata, me te pānui tūtohi hei kaha o te whānau, kaua hei whakaaro-tōmuri.

MiniMax-M1 — te aro whakaranu mō te whakaaroaro roa

I whakaahuatia a MiniMax-M1 (arXiv 2506.13585) e ōna kaituhi hei tauira whakaaroaro aro-whakaranu rahi-nui, taumaha-tuwhera tuatahi. I pupuri tonu i te tinana MoE 456B / 45.9B, ā, i whakawhitiwhiti i ngā momo aro e rua: he poraka transformer aro-softmax māori i muri i ia whitu poraka Lightning Attention. I inea te painga i te tatau, kaua i te tūnga papatuhi-toa — i pūrongo a MiniMax i te roa whakaputa 100K-tohu, i whakapau a M1 i te tata 25% o ngā FLOPs a DeepSeek R1, i te wā e tautoko taketake ana i tētahi horopaki kotahi-miriona-tohu.

MiniMax-M2 — te huri hohenga-iti

I haere a MiniMax-M2 ki te ahunga kē mō te rahi, ā, i toa i te aro mō tērā: he kaiwhakawaehere-anō 62-paparanga me te 229.9B tawhā katoa engari ~9.8B anake e hohe ana ia tohu, e whakamahi ana i te 256 mātanga matatini ia paparanga MoE me te whakarite-ara sigmoid ki te waru o rātou, ngā tūranga RoPE me te QK-RMSNorm mō te pūmau whakangungu. He tauira waehere-kaihautū mārōrō tāna — ngā whakatikatika kōnae-maha, ngā porowhita rere-whakatika, ngā whakatikatika kua whakaūngia-whakamātautau — ā, i whakaurua e ia te hōputu whakaaro whakawhitiwhiti, i reira ka kawea te whakaaroaro i waenga i ngā porowhita karanga-taputapu, kaua e whakarerea i muri i ia hurihanga. Ko M2 te tauira nāna i whakatū te rongonui o MiniMax mō te tuku whanonga kaihautū paetata-ā-paetahi mō tētahi wāhanga o te utu ia-tohu, mā te pupuri noa i ngā hohenga iti.

MiniMax-M3 — te tūāpapa kei raro i tēnei kūwaha

I whakaputaina a MiniMax-M3 i te 1 o Hune 2026, me ngā taumaha ki Hugging Face i muri tata iho, me tētahi pūrongo hangarau i te arXiv 2606.13392. He tauira Whakaranu-o-Ngā-Mātanga 428B-tawhā me te tata 23B hohe ia tohu, he horopaki kotahi-miriona-tohu kua tuhia, me te tāuru whakaahua me te ataata taketake — e whakaahua ana a MiniMax i a ia he mea whakangungua reo-maha-whakaranu "mai i te hipanga tuatahi rawa," kaua he mea kuputuhi-whakangungua-tōmua kātahi ka whakaritea reo-maha. I runga i ngā tau a MiniMax ake i whakaputaina, ka eke ia ki te 80.5% i te SWE-bench Verified me te 59.0% i te SWE-bench Pro. He tau nā te kaihoko i pūrongotia mō tētahi paearu waehere ēnei, ā, koinā tonu te āhua e tika ana kia taumahatia e koe; ka tohuhia e mātou nā te mea koinā ngā kerēme i whakaputaina e te whare rangahau, kaua nā te mea kua whakatauīra motuhaketia e mātou.

03MiniMax Sparse Attention: he aha i taea ai e tētahi tauira tirohanga te iti tēnei utu

Ko te hangarautanga whakamīharo i M3 ko te MiniMax Sparse Attention (MSA), ā, koinā te take tika ka taea ai e tētahi tauira reo-maha o tēnei rahi te noho i tō mātou papa utu. He utu-nui ngā tohu mō ngā whakaahua. Ka taea e ētahi whārangi iarere-teitei te whakauru i ngā tekau mano tohu ki tētahi tohutohu i mua i tō tuhinga i tētahi tohutohu kotahi. I raro i te aro paerewa, ka whakarea whā te utu, ā, ehara te tauira e pānui ana i ngā whārangi rua tekau i te rua whakarea te utu-nui i te mea e pānui ana i te tekau — he kino ake i te rua whakarea.

Ka panoni te MSA i te āhua o taua kōpiko. He tikanga poraka-korekore kua hangaia i runga i te tuaiwi Grouped Query Attention, kua wehea ki ngā manga e rua:

• He Manga Tohu (Index Branch) māmā e whakatau ana i ngā poraka kī-uara — ngā tatauranga max-pooling puta noa i ngā tohu o roto i ia poraka — ā, ka kōwhiri i te top-k mō ia tūranga pātai me ia rōpū GQA, e whakauru tonu ana i te poraka ā-rohe e mau ana i te pātai ake mō te pūmau whakangungu.
• He Manga Matua (Main Branch) e whakahaere ana i te aro softmax māori ki ngā poraka anake i kōwhiritia e te tohu.

Ko te whirihoranga i te pepa MSA he k = 16 poraka o te 128 tohu, e homai ana ki ia pātai he kōti pūmau 2,048 tohu aro. He pūmau, arā kāore e tipu me te roa o te raupapa. Koirā katoa te hangarau: ka tipu haere te horopaki, ka piki tonu te utu aro o tētahi paerewa kiato, engari ka noho papatairite te utu manga-matua a MSA, nō reira ka tipu ōrite te aro, kaua e whakarea whā.

He tatauranga nō te pūrongo MiniMax Sparse Attention (arXiv 2606.13392): he 28.4× te heke o ngā FLOPs aro ia-tohu hei whakataurite ki te GQA whirihoranga ōrite i te horopaki 1M, me te 14.2× te tere prefill / 7.6× te tere decode i inea i runga i ngā taputapu H800 i te rangahau 109B-tawhā a te pepa. E pūrongo motuhake ana a MiniMax i te 9× te tere prefill me te 15× te tere decode mō M3 hei whakataurite ki a M2 i te horopaki 1M. He mea nā te whare rangahau katoa i pūrongotia.

Ka mahi hoki te pepa i te tirotiro e whai take ana mō tētahi tuhinga tirohanga: i runga i tētahi MoE 109B reo-maha taketake, i mau te MSA i tētahi mahi e rite ana ki tētahi paerewa aro-katoa puta noa i ngā paearu whakaahua tae atu ki te AI2D, ChartQA me te MMMU, me ngā paearu ataata tae atu ki te EgoSchema, LongVideoBench me te VideoMME — me ngā momo korekore i ētahi wā e whakapai ake ana i ngā mahi ataata horopaki-roa ake. Ehara te korekoretanga i konei i te taake kounga e utua ana e koe mō te tere; i ngā tāuru roa, he tata kore-utu tōna āhua.

He aha koe e whai take ai hei kaihoko: ko te utu whakahaere te mea e whakatakoto ana i te utu rārangi. Ko tētahi tūāpapa kua mutu te tipu o tōna utu aro he tūāpapa ka taea e koe te whakamahi ki te rau mano whakaahua. Koinā te tikanga kei muri i te "tauira tirohanga kore-ārai utu-iti rawa" — ehara i te reiti panui, he hoahanga.

04He aha te tikanga o te 3BIT-REAP, ā, kei hea e utu ana koe

Kāore tētahi tauira 428-piriona-tawhā e uru māmā ki tētahi wāhi i te tika katoa. Kei tā mātou hanga MiniMax-M3 te taunga 3BIT-REAP, e whakaahua ana i ngā mahi motuhake e rua i whakamahia i tētahi raupapa tūturu.

REAP: mukua ngā mātanga i mua i tō kōpaki i a rātou

Ko te REAP — Router-weighted Expert Activation Pruning — he tikanga kōpaki MoE hurihanga-kotahi nā Cerebras me te University of Calgary, i whakaputaina hei arXiv 2510.13999, i whakaaetia mō ICLR 2026. Ko tāna tautohe he pai ake te tapahi mātanga i te whakakotahi mō ngā mahi whakaputa: ka whakangaro te whakakotahi i te mana matatini o te kaiwhakarite ara, ā, ka whakauru i tētahi hapa kāore koe e taea te whakahoki mai. Nō reira, ka whakatau kē te REAP i ia mātanga mā tētahi paearu whai-take e whakakotahi ana i ngā uara kūaha kaiwhakarite ara (te auau me te kaha o te kōwhiritanga a te kaiwhakarite ara) me ngā paerewa hohenga mātanga (te nui o tāna whai wāhi whakaputa tūturu), kātahi ka muku tuatahi i ngā mātanga tatauranga-iti rawa.

E pūrongo ana ngā kaituhi i te muku tae atu ki te 50% o ngā mātanga mai i ngā tauira tae atu ki te 1 taera tawhā me te pupuri tonu i te kounga paerewa — i runga i te Qwen3-480B-Coder-FP8 i te 50% tapahi, i mau tonu te 97.6% o te āheinga waehere ehara-i-te-kaihautū paerewa me te 96.7% i te SWE-Bench kaihautū. He tau ēnei nā te pepa mō ngā tauira a te pepa, ehara i a mātou mō MiniMax-M3; ka tohuhia e mātou kia taea e koe te whakatau i te tikanga, kaua hei tohu he ine ōrite kei konei.

Ngā taumaha 3-bit, tuarua

Ka rongoātia ngā taumaha e ora tonu ana ki te 3 pae. He mea ata whakaaro te raupapa, ā, koinā te take o te whakakotahitanga: tapahi tuatahi, quantize tuarua. Kāore he tikanga ki te whakapau i ō toru pae mīharo mō ia taumaha ki te whakawaehere pono i ngā mātanga e tata mukua ana e koe. Tapahia te hiku korekore, kātahi tukua te kōti e toe ana kia haere katoa ki ngā mātanga e tino hohe ana.

Te whakatūpato pono

Ko te 3BIT-REAP tā mātou taumata kōpaki kaha rawa. Ka whakamahia e mātou i ngā tūāpapa nunui rawa — MiniMax-M3, Kimi-K3, DeepSeek-V4-Pro, GLM-5.2, Nemotron3Ultra — nā te mea koirā te rerekētanga i waenga i te wātea o aua tauira i tētahi utu tika me te kore wātea rawa. I te wāhi he iti noa te tūāpapa kāore e hiahia, kāore mātou e whakamahi: ka whakahaere ngā tauira W8A16 o te rārangi i ngā taumaha 8-bit me ngā hohenga 16-bit, ā, e noho ana te W4A16-AUTOROUND-REAP i waenganui, e whakamahi ana i te porowhita kua akona a AutoRound ki te whakahoki mai i te tika ka tukua e te quantization porowhita-ki-te-tata-rawa māori. Kōwhiria te taumata e rite ana ki tō manawanui. Mēnā kei te tango koe i ngā āpure mai i te tekau mano nama-tāke, koinā te whakawhitinga tika te 3-bit. Mēnā kei te mahi koe i tētahi pānui taihara tūraru-teitei kotahi mō tētahi whakaahua kotahi, whakapaua te moni ki a Kimi-K3.

05He aha te tikanga o "kāore he paparanga whakakāhore" ina he whakaahua te tāuru

He kupu te "kore-ārai" e whakamahia māmātia ana, nō reira anei te kerēme tino tika: kāore he paparanga whakakāhore, kāore hoki he tātari whakaputa i tēnei kūwaha. Kāore tētahi kaiwehewehe e tirotiro ana i tō whakaahua i mua i te kite a te tauira. Kāore tētahi tauira kaupapa here e tirotiro ana i ngā tohu i tōna putanga. Ko tāu ka riro mai ko te mea i hangaia e te tauira.

I te kuputuhi, he mōhiotia te pānga o tērā. I te tirohanga, he rerekē tōna momo, nā te mea ko ngā tāuru whakaahua te wāhi kaha rawa te taupatupatu me te noa-noa o te whanonga whakakāhore i ngā API reo-maha arumoni. Ko ngā hapa āhua e whai ake nei:

• Ngā whakakāhoretanga tuakiri whānui — te whakakāhore rawa ki te whakaahua i ngā tāngata i tētahi whakaahua, tae atu ki ngā āhuatanga e kitea mārama ana, e kore rawa e mōrearea, nā te mea kāore e taea e tētahi kaupapa here te wehe i te whakaahuatanga mai i te tautuhi tuakiri.
• Ngā omanga kaupapa — te whakakāhore ki te tuhi anō, te whakarāpopoto rānei i tētahi tuhinga nā te take o te tuhinga, ahakoa he mahi miihini te tuhi anō.
• Te mataku rongoā me te taihara — te whakakāhore ki te whakaahua i ngā whakaahua haumanu, ngā tuhinga whara, ngā whakaahua aitua e hiahiatia ana e tētahi mataaho kia whakaahuatia.
• Ngā tika-hē o te taputapu haumarutanga — te whakakāhore ki te pānui i tētahi hopukanga mata nā te mea kei roto he rārangi whakahau, he hopu pūkete, he kaiwhakatikatika hapa rānei.
• Te heke ngū — te mea kino rawa: ehara i te whakakāhoretanga engari he whakautu-kore kua parenga, e paahitia ana i tō whakaūnga hanganga, ā, e whakapaitini ana i tō huinga raraunga me te kore rawa e ara ake he hapa.

Ko te raru whakahaere o aua mea e rima, ehara i te kore-tika. Engari he kore-ine. He tohu raraunga tētahi whakautu hē; ka taea e koe te tatau, te hopu i te arotake, te whakatika. He rua kau tētahi whakakāhoretanga. I te taumata pupuri, ko tētahi reiti whakakāhoretanga 3% e tohatohaina ana kore-matapōkere puta noa i āu tāuru — e whakakohia ana, mā te hanga, ki ngā take pakeke me te onoono tonu — he whakaaro-tuara pūnaha i ō hua kāore he penapena o muri e whakahoki mai ai. Ko te tango i te paparanga whakakāhore te tango i tētahi puna raraunga ngaro ngū, honohono.

Ka whakahaerea a Shannon e Shannon Lab LLC, New Mexico, USA, i raro i tā mātou Kaupapa Here Whakamahi Haepapa. Ko te tikanga o te kore-ārai kāore te tauira e whakaputa whakaaro mō āu tāuru; ehara i te tikanga kāore he tikanga o te pūkete.

06Te tautuhinga katoa

Wāhanga	MiniMax-M3 · 3BIT-REAP
Id tauira	MiniMax-M3-3BIT-REAP
Tūāpapa	MiniMax-M3 (MiniMax AI)
Matapihi horopaki	262,144 tohu
Whakaputa nui rawa	16,384 tohu
Tāuru whakaahua	Āe
Tāuru ataata	Kāore e whakaaturia
Streaming	Āe
Karanga taputapu	Āe
Whakaaroaro	Āe (reasoning_effort)
Putanga hanganga	json_schema + response_format
Paparanga whakakāhore	Kāore
Ngā kūwaha	/v1/chat/completions · /v1/messages · /v1/responses
E whakahaerea ana i runga i	Tā Shannon ake tautau GPU

E rua ngā rārangi o taua ripanga e tika ana kia whakanuia. Kāore te tāuru ataata e whakaaturia. I whakangungua te tūāpapa M3 pito-tōmua ki te ataata, ā, e tuhi ana a MiniMax i te tāuru ataata; ka whakaae tā mātou kūwaha ki ngā whakaahua anake. Hangaia ki ngā whakaahua. Tuarua, nō mātou te matapihi 262,144-tohu, kaua nō te tau pito-tōmua kotahi-miriona — whakamaheretia ngā kawenga mahi tuhinga-roa me te whakaahua-maha ki te 262,144, ā, kāore koe e whakamīharotia.

07Te karanga i te API MiniMax kore-ārai

E toru ngā reo, ko te tauira ōrite kei muri i a rātou katoa, ka streaming i ngā mea e toru. Whakamahia te mea e kōrero kē ana tō kiritaki; kāore he rerekētanga āheinga i waenga i a rātou. Tohua tō URL tūāpapa ki a Shannon, ka huri i te id tauira.

Te māramatanga whakaahua mā te reo 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
  }'

Ka mahi ngā URL https:// mamao me ngā URL raraunga base64 i te wāhanga image_url. Whakapiritia ētahi wāhanga whakaahua ki te karere kotahi ina hiahia koe i te whakataurite whārangi-ki-te-whārangi — he 262,144 tohu tāu hei whakapau, ā, i te $0.50 ia miriona, he utu-iti te whakapau i a rātou.

Te tango kua whakatohu mā te JSON schema

Ko te wāhi e utu ai tēnei tauira i tōna tūranga i tētahi ara ko te whakakotahitanga o te tirohanga + hanganga pūmau + papa utu. Ehara i te mea kei ngā tauira katoa o tā mātou rārangi ēnei mea e toru: kāore a Laguna-S-2.1, inkling rānei e tautoko i te json_schema. Ka tautoko a MiniMax-M3, ko te tikanga ka taea e koe te whiwhi i tētahi ahanoa kua whakaūngia mai i tētahi whakaahua me te kore paparanga wetewete:

{
  "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
      }
    }
  }
}

Te kaha whakaaroaro

Ka whakaputa te tauira i te reasoning_effort. Mō te whakarōpū whakaahua pukuputu, waiho iti — kei te pātai koe i tētahi pātai āhua-ā-kanohi, ā, he tohu whakaputa ka utua e koe ngā tohu whakaaroaro. Whakanuia mō ngā karanga i reira te whakaahua he taunakitanga i tētahi tautohenga roa ake: te pānui i tētahi hoahoa e whakataurite ana ki tētahi tautuhinga, te whakahoahoa i tētahi tūtohi ki tētahi ripanga, te whakatau mēnā e rua ngā hopukanga mata e whakaatu ana i te hapa ōrite.

Kei te tuhinga API a Shannon te tohutoro tawhā katoa, te tikanga hapa, me ngā hōputu streaming mō ngā reo e toru.

08Kei hea te tika o MiniMax-M3 — kei hea kāore

Toro atu ina

• Te mahi whakaahua ranea. Te tapanga huinga raraunga, te rangahau whakahaere-ihirangi, te whakarawaka rārangi, te tirotiro tuatahi hopukanga mata, te OCR-me-hanganga pukuputu i ngā tuhinga kua matawaihia. Ko te papa utu te tautohe katoa.
• Ngā whakaahua ki ngā rekoata kua whakatohu. He whakakotahitanga e rua anake ngā tauira o konei e tuku ana: te tirohanga me te json_schema pūmau me ngā tohu utu-iti rawa o te rārangi.
• Te horopaki reo-maha roa. He maha ngā whārangi, he maha ngā papanga rānei, i tētahi tono kotahi — kua hangaia te hoahanga o raro mō tēnei tonu āhua tāuru.
• Te rangahau ka whakakinotia e ngā whakakāhoretanga te tauira. Ia rangahau e whakawhirinaki ana tōna pono ki te whakautu a te tauira i ia tūemi, tae atu ki ngā mea kāore e pai.
• Ngā porowhita kaihautū e pānui ana i tētahi mata. He mārōrō kaihautū te whakapapa raupapa-M, ā, kei te wātea te karanga taputapu.

Toro atu ki tētahi atu ina

• Kei te hiahia koe i te tuanui tirohanga teitei rawa, ā, kei te wātea te pūtea. He take tā Kimi-K3 · 3BIT-REAP mō tōna 7.7× te utu tāuru; mō tētahi pānui tūraru-teitei kotahi, koinā te moni tika hei whakapau.
• Kei te hiahia koe i ngā taumaha pono-rawa. Ko te 3-bit tā mātou taumata kōpaki kaha rawa. Ka whakawhiti ngā tauira W8A16 i te utu mō te tika, ā, he tirohanga tā MiMo-V2.5 · W8A16 i te $0.50 / $2.00 ōrite mēnā ko te pono ia-taumaha tāu e whakapai ana.
• Kei te hiahia koe i te ataata. Kāore mātou e whakaatu i te tāuru ataata i tēnei kūwaha.
• He kuputuhi anake, taumata paetata, te mahi. Kei te noho a DeepSeek-V4-Pro-0813 mō tērā; kei te utu koe mō tētahi pou tirohanga kāore koe e whakamahi.

He tatauranga hoahanga, paearu me te tere-piki mō te whānau MiniMax i pūrongotia e MiniMax AI; he tatauranga REAP i pūrongotia e ngā kaituhi REAP mō ā rātou ake tauira, ehara i te ine mō tēnei hanga. Nā mātou ngā utu, te matapihi horopaki me ngā haki āheinga a Shannon, ā, ka panuitia mai i tā mātou rārangi tauira ora.

09Ngā pātai auau

---

Puna kōrero: Kāri tauira 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. He tatauranga paearu me te tere-piki nā MiniMax i pūrongotia, ā, he mea whakaharatau tonu i te tatari ki te whakaū motuhake. He mea ora ngā utu me ngā haki āheinga a Shannon, e rite ana ki te 5 o Hepetema 2026.