API Tauira Inkling Kore-ārai
He hanga whai-tirohanga, kore-ārai o te tūāpapa Whakaranu-o-Ngā-Mātanga 975B Inkling a Thinking Machines — horopaki 262K, tāuru whakaahua, kore paparanga whakakāhore.
Whakarāpopoto
Ko inkling-W4A16-AUTOROUND-REAP tā mātou whakatakotoranga o Inkling, te MoE taumaha-tuwhera 975B-tawhā i whakaputaina e Thinking Machines Lab i te 15 o Hurae 2026. He whai-tirohanga, ka mahi i te karanga taputapu me te whakahaere kaha whakaaroaro, ā, ka rere me te kore paparanga whakakāhore, kore hoki tātari whakaputa i runga i tā mātou ake tautau GPU. Ko te horopaki he 262,144 tohu, ko te whakaputa tae atu ki te 16,384. Ko te utu he $1.42 tāuru / $6.07 whakaputa ia miriona tohu — waenga-nui-ake, kei runga ake i tā mātou taumata papa $0.50, ā, kei raro noa iho i te tauira tino teitei a Kimi-K3. Kotahi whakatūpato hei whakamahere: ka tautoko ia i te response_format engari kāore i te json_schema.
Ko te nuinga o ngā tauira tekau mā rua o tā mātou kūwaha he kuputuhi-anake. E whā kāore i pērā, ā, i waenga i aua tauira e whā, he tūranga motuhake tō Inkling: te tauira tirohanga ka tono koe ina ko te whakaahua te wāhanga uaua o te rapanga, engari kāore te pūtea e roa ki tētahi tauira tino teitei. Ka kōrero tēnei kāri mō te ahunga mai o te tūāpapa, te mea i mahia e te taunga W4A16-AUTOROUND-REAP ki a ia, ngā mea ka mahia me ngā mea kāore e mahia i te mata API, ā — kia pono — te wā he pai ake te karanga i tētahi tauira utu-iti ake o te rārangi ōrite.
01He aha a Inkling, ā, nā wai i hanga?
Ko Inkling te tauira taumaha-tuwhera tuatahi a Thinking Machines Lab, i whakaputaina i te 15 o Hurae 2026. Kāore i rite ki te nuinga o ngā putanga e kīia ana he "tuwhera," i tukuna ia me ngā taumaha tāngohia-katoa i runga i a Hugging Face me te tautoko-rā-tuatahi i roto i a transformers, SGLang, vLLM, me llama.cpp — koirā tonu te take ka taea ai te whakahaere i runga i te hanganga pēnei i tā mātou. Ngā tau matua, mai i ngā rauemi putanga ake a Thinking Machines:
He whakatakotoranga Whakaranu-o-Ngā-Mātanga korekore tēnā: 66 paparanga wetereo i te rahi huna 6,144, ko ngā mea tuatahi e rua he kiato, ko ngā 64 e toe ana e kawe ana i te 256 mātanga kua whakaritea-ara me te 2 mātanga tiritahi, me te 6 mātanga kua whakaritea-ara e hohe ana ia tohu. Tata ki te 4% o ngā tawhā e mahi ana i tētahi tohu; koirā te mea e taea ai te whakahaere i tētahi tauira tata-taera-tawhā.
E rua ngā kōwhiringa hoahanga he tino rerekē kia whakaingoatia. He aro whakawhitiwhiti: 55 paparanga aro matapihi-neke i tētahi matapihi 512-tohu, e whakataurite ana ki te 11 paparanga aro-whānui, i tētahi tauira tukurua 5:1. Ā, kāore a Inkling e whakamahi i te RoPE: ka whakamahia e ia he whakapēhitanga tūnga-whanaunga kua akona, ā-tāuru, me ngā convolution poto take-tōmua i runga i ngā projection kī/uara me ngā peka toenga. E tohu ana ngā tuhipoka hoahanga a Sebastian Raschka i ēnei e rua hei nekehanga tūturu mai i te tohutaka paerewa o te tau 2026. I whakamahia te Muon mō ngā tawhā matriki nunui i te whakangungu, me te Adam mō ērā atu, kātahi neke atu i te 30 miriona rere RL kore-rite-wā i te whakangungu-whai-muri.
Ehara te take ki tētahi kaiwhakamahi API i ngā mea iti noa. Koinā he tūāpapa rahi-taumata-mata, maha-momo ā-taketake, i whakangungua ki te 45 taera tohu e whātoro ana i te kuputuhi, ngā whakaahua, ngā oro me ngā ataata — kaua he tauira kuputuhi me tētahi taputanga tirohanga i whakapiritia mai i muri. Ka whakawaeahia ngā whakaahua hei wāhanga pika 40×40 mā roto i tētahi kaiwāhi MLP whakarōpū e whā paparanga, ka tukatukaina ngātahitia me te kuputuhi i te wetereo ōrite. I roto i te whāinga whakangungu-tōmua te mārama-ataata.
02He aha te tikanga o W4A16-AUTOROUND-REAP?
E toru ngā tikanga motuhake kua whakapaparuatia i taua taunga, ā, he rerekē ngā mahi a tēnā, a tēnā. He mea whai take kia mārama, nā te mea koinā te rerekētanga nui rawa i waenga i te tauira ka karangatia e koe i runga i tā mātou API me te pauna i whakaputaina e Thinking Machines.
REAP — te tapahi i ngā mātanga kāore e whakaritea-ara e tētahi
Ko REAP te Router-weighted Expert Activation Pruning, nā Cerebras Research (arXiv 2510.13999, kua whakaaetia mō ICLR 2026). I roto i tētahi MoE me te 256 mātanga ia paparanga, kāore te kaiwhakarite ara e tohatoha ōrite ana i te mahi: he auau te pā ki ētahi mātanga, ā, he onge te puta o ētahi, he iti hoki tā rātou whai wāhi ina puta. Ka whakatau te REAP i ia mātanga i runga i ngā tohu e rua — ngā uara kēti a te kaiwhakarite ara me ngā ine whakaputa a te mātanga — he iti ake te take o tētahi mātanga e auau ana te kōwhiritia engari he iti tāna whai wāhi i te mea e kīia ana e tōna tatauranga kōwhiringa — ā, ka tukuna ngā mea tatauranga-iti rawa i te hikoi kotahi, i mua i te quantization.
Ko te take ka whakamahia e mātou te tapahi kaua ko te whakakotahi mātanga rongonui ake, he rerekē ngā mea e rua i runga i te momo mahi. He tino pai te āhua o te whakakotahi i ngā paearu wehewehe, engari ka horo i ngā mea whakaputa. E pūrongo ana ngā kaituhi o REAP he pai ake te pupuri kounga a te tapahi i ngā kawenga mahi tonu e hokona ai e te tangata tētahi API — te whakahiato waehere, te karanga taputapu, te whakaaroaro maha-hurihanga — i ngā ōwehenga kōpaki neke atu i te 50%, puta noa i ngā tauira mai i te 20B ki te 1T tawhā. Ka whakangaro te whakakotahi i te mātau-motuhake a te kaiwhakarite ara; ka waiho e te tapahi ngā mātanga e ora tonu ana me te whakaritea-ara i waenga i a rātou.
AutoRound — te ako ki hea te porowhita
Ko AutoRound te tikanga quantization-whai-muri-whakangungu a Intel (arXiv 2309.05516, i whakaputaina hei SignRound). Ka porowhita te quantization ngāwari i ia taumaha ki te uara 4-pī tata rawa — he taunoa whaitake engari he kino i te rahi, nā te mea ehara ngā hapa porowhita i te mea motuhake, ā, ka whakarea puta noa i te paparua 66-paparanga. Ka kite a AutoRound i te whakatau porowhita hei mea hei whakatikatika: ka whakatikatika ngātahi te heke-paiaka-tohu i te ahunga porowhita me ngā awhe kutikuti taumaha puta noa i te rua rau hikoi, ā, ka nama i te tika o te whakangungu mataara-quantization me te kore whakangungu-anō. E pūrongo ana a Intel i ngā piki nui rawa i ngā tautuhinga pī-iti rawa, me tētahi tuanui whakataetae i te INT4.
W4A16 — kei hea ngā pī e haere ana
Ko W4A16 te hōputu tatauranga: ngā taumaha 4-pī, ngā hohenga 16-pī. Ko ngā taumaha — e nui rawa ana i te mahara o tētahi tauira 975B — ka whakaitihia ki te tata hauwhā o tētahi horahanga BF16, engari ka noho tonu te tatauranga ki te 16 pī. He nui taua wehenga mō tētahi tauira tirohanga ake: ka pāngia e ngā tikanga hohenga-quantized tonu te mahi tatauranga mārama-iti e whakawhirinaki ana te mārama whakaahua ki a ia — te pānui i tētahi tapanga rārangi ine iti, te whakatau i tētahi rārangi angiangi i roto i tētahi hoahoa, te wehewehe i ngā tae rite e rua. Ko te pupuri i ngā hohenga ki te 16 pī he utu ki te whānui-rārangi, engari ka hoki mai te tika ki ngā wāhi e mōhiotia ana e tētahi kawenga mahi tirohanga.
Mā te whakapaparua i ngā mea e toru, ka whiwhi i tētahi tauira e uru ana, e tere tonu ana te rere, me te kore paremata kounga o te taumata 3BIT-REAP kaha ake e whakamahia ana e mātou mō ngā tūāpapa nui rawa. Koirā te whakawhitinga e whakaatahia ana e te utu.
03Te tirohanga: he API tātaritanga whakaahua kore-whakakāhore
Ko Inkling tētahi o ngā tauira e whā e whai-tirohanga ana i tā mātou rārangi, i te taha o Kimi-K3, Kimi-K2.6, me MiniMax-M3. Ko ngā hua tirohanga i whakaputaina o te tauira taketake, mai i te putanga a Thinking Machines, he 73.3% i te MMMU Pro (Paerewa 10) me te 78.1% i ngā pātai whakaaroaro CharXiv — ka piki ki te 82.0% ina whakaaetia te tauira ki te whakahaere Python ki te tūtohi. He tau nā te whare rangahau anō ērā mō te pauna kāore i kōpakina; tirohia hei tuanui o te whānau, kaua hei ine mō tā mātou hanga.
Ko CharXiv te mea whai take kia āta tirohia, nā te mea e whakaatu ana i te take ka hoko ai te tangata i tētahi API tirohanga, kaua i tētahi ratonga OCR. Ka whakamātauria e ia te whakaaroaro i runga i ngā tūtohi pūtaiao — te pānui i tētahi uara mai i tētahi rārangi ine, te whakataurite i ngā raupapa e rua, te kite i te hurihanga o tētahi ia. Ehara i te "he aha te kuputuhi kei roto i tēnei whakaahua" engari he "he aha te kī a tēnei whakaahua, ā, e mau ana rānei taua kī." Koinā te mea e whakaū ana i te tuku whakaahua ki tētahi tauira whakaaroaro 975B, kaua ki tētahi kaiwhakamārama iti.
He mahi noa rite-ki-OpenAI te tuku whakaahua — he wāhanga ihirangi image_url i runga i /v1/chat/completions, he poraka image momo-Anthropic i runga i /v1/messages, he ōrite rānei i runga i /v1/responses. Ka pā ngā reo e toru katoa ki ngā taumaha ōrite, ā, ka streaming katoa; kōwhiria te mea e kōrero ana tō SDK o nāianei. Kei te wāhanga 09 tētahi tauira mahi.
He nui ake te take o te kore-ārai mō te tirohanga i te kuputuhi. He tino māro tētahi paparanga whakakāhore i runga i tētahi pito mutunga tirohanga, nā te mea ka whakarōpūhia paerewatia te whakaahua i mua i te whakaaroaro a te tauira mō te tono. Ka mau tonu te mahi tika: he whakaahua whara i roto i tētahi paipa tuhinga haumanu, he whakaahua-mata taunakitanga i roto i tētahi tuhinga aituā, he papa mai i tētahi kiriata e arotakengia ana mō tētahi tuhinga arotake, he mataono rongoā, he whakaahua o tētahi tuhinga e mau ana i tētahi ingoa. Kāore he kaiwhakarōpū pēnā i mua i a Inkling i runga i tā mātou kūwaha. Ka tirotiro ia i te mea i tukuna e koe, ka whakautu i te pātai i pātaihia e koe. Koinā te hua.
04Putanga hanganga: response_format āe, json_schema kāo
Koinei te wāhi kotahi he iti ake te āheinga o Inkling i te nuinga o te rārangi, ā, he pai ake te whakamahere mō tēnā i te kite noa i te wā whakaputa.
| Āheinga | inkling | Kimi-K2.6 | MiMo-V2.5 | Laguna-S-2.1 |
|---|---|---|---|---|
| Tirohanga (tāuru whakaahua) | Āe | Āe | Āe | Kāo |
response_format | Āe | Āe | Āe | Kāo |
json_schema | Kāo | Āe | Āe | Kāo |
| Karanga taputapu | Āe | Āe | Āe | Āe |
| Kaha whakaaroaro | Āe | Āe | Āe | Āe |
Ka mahi te response_format: {"type": "json_object"} — ka whiwhi koe i te JSON tika-ā-wetereo. Ko te mea kāore koe e whiwhi ko te wetewete kua herea-ā-wetereo ki tētahi schema kua tukuna e koe, i reira kua aukatingia ā-tinana te kaitango i te whakaputa i tētahi tohu e takahi ana i tō hanganga. Ko ngā ingoa āpure, te mema-enum, me te noho o ngā kī matua he mea nā te tauira ake, ehara i te mautohu.
E toru ngā ara ki te noho tahi me tēnā, kua raupapatia i runga i te pono:
- Tonoa te schema, ka whakaū. Whakaurua te schema ki te tohutohu pūnaha hei JSON tūturu, whakaritea te response_format ki json_object, ka whakaū i te whakautu i tō taha. Tarai anō ina hapa te whakaū, me te tāpiri i te karere hapa a te kaiwhakaū. He utu-iti, ka mahi, engari me whai pūtea-tarai anō.
- Whakamahia he whakamārama taputapu hei schema. Ka tautokona te karanga taputapu, ā, ko te whakamārama tawhā a tētahi mahi he JSON Schema. Whakaingoatia he taputapu kotahi mō te tango e hiahiatia ana, akiakitia, ka pānui i ngā tohutohu. Koinei te kōwhiringa kaha rawa i runga i a Inkling i te nuinga o te wā, ā, kāore he utu tāpiri.
- Wāhia te mahi ki ngā tauira e rua. Waiho a Inkling kia mahi i te tirohanga me te whakaaroaro hei kuputuhi, kātahi ka hoatu taua kuputuhi ki a
MiMo-V2.5-W8A16i te $0.50 / $2.00 me tētahijson_schematūturu mō te hikoi whakahanga. E rua ngā karanga, engari he utu-iti te tuarua, he mautohu te āhua, ā, ko te tauira e hiahiatia ana kia mātau ko te mea i kite i te whakaahua.
Ko te kōwhiringa 3 te tauira e taunaki ai mātou mō tētahi paipa tango-kōrero rahi-nui. Utua te utu nui ake mō te hikoi anake e hiahia ana ki tēnā.
05He aha ka hokona e te utu nui ake?
I te $1.42 tāuru / $6.07 whakaputa ia miriona tohu, e noho ana a Inkling ā-hiahia ki waenganui o tā mātou rārangi. E ono o ngā tauira tekau mā rua kei te papa $0.50 / $2.00. Ko Kimi-K3 — te tauira tino teitei, he whai-tirohanga hoki — he $3.83 / $19.12. Ko Inkling tata ki te toru whakareatanga o te papa mō te whakaputa, tata hoki te hautoru o te tauira tino teitei.
Utu whakaputa, USD ia 1M tohu, ngā reiti rārangi a te Kūwaha Shannon AI, Hepetema 2026. He rite te whānui pae ki te utu; he roa ake he utu nui ake.
Anei te pānui pono mō te wāhi e whai utu ana taua utu nui ake — me te wāhi kāore.
| Kawenga mahi | Whakamahia a Inkling? | He aha |
|---|---|---|
| Whakaaroaro tūtohi / hoahoa | Āe | Ko ngā tau CharXiv me MMMU Pro i whakaputaina te mea e kaha rawa ai tēnei tūāpapa. |
| Ngā kaihautū whakaahua-mata | Āe | He tirohanga, he karanga taputapu, he 262K hītori kei te tauira kotahi, kāore he tuku-atu. |
| Ngā tuhinga roa me ngā whakaahua | Āe | Ka tiritahi te kuputuhi me ngā whakaahua i te wetereo kotahi me te matapihi horopaki kotahi. |
| Te hanga kuputuhi noa | Kāo | E ono ngā tauira i te $0.50 / $2.00 e mahi ana i tēnei. Kei te utu koe mō tētahi ara tirohanga kāore koe e karanga. |
| Te tango schema pukoro | Kāo | Kāore he json_schema. He utu-iti ake, he here ake hoki a MiMo-V2.5. |
| Te tuanui mata tino teitei | Kāo | Ko Kimi-K3 te tauira tino teitei. Utua te 3.1× mēnā ko te tuanui te here. |
| Te tirohanga utu-iti rawa | Kāo | He tirohanga tō MiniMax-M3 i te $0.50 / $2.00. Tīmata i reira, whakanekehia mēnā ka akiaki te kounga. |
Pānuitia tēnā hei ture whakarite-ara, kaua hei rārangi tūranga. He whāiti ake te horopaki mahi e tika ai a Inkling i te mea ka kīia e tētahi whārangi hokohoko, ā, i roto i taua horopaki, he uaua te aha ake i tōna utu. I waho atu, ka pai ake ki a mātou te hoko atu i te tauira utu-iti ake ki a koe, ā, kia noho tonu koe hei kaihoko.
06Te matapihi horopaki 262K, ā, he aha tino ka uru
Ka whakahaeretia ngā tauira tekau mā rua katoa o te kūwaha me te horopaki tāuru 262,144-tohu me te whakaputa nui rawa 16,384-tohu ia tono. He mea ōrite ā-hiahia: kaua rawa te whakahaere horopaki e noho hei peka ā-tauira i roto i tō waehere, ā, ko te whakakapi i a inkling ki a MiniMax-M3 i roto i tētahi kōnae whirihoranga kaua e panoni pūku i te mea ka uru.
Kia mōhio ki te āputa ki te tūāpapa i whakaputaina: ka tuhia e Thinking Machines a Inkling ki te 1,048,576 tohu, ā, ka whakahaerea e mātou he wāhi 262K o tēnā. Mēnā he kawenga mahi tono-kotahi tohu-miriona tūturu tāu, ehara mātou i te pito mutunga mō tēnā, ā, he pai ake kia mōhio koe i mua i tō hanga.
Mō te tauine i te 262K: he PDF hangarau 400-whārangi me ōna whakaahua, he pātengi waehere rahi-waenga, ētahi haora tuhinga-kōrero, he wātea kaihautū roa e kawe ana i ngā tekau hua taputapu me te ringa whakaahua-mata — ka uru ia tētahi ki te tono kotahi. I runga i tētahi tauira tirohanga, kia maumahara ka pau anō te horopaki e ngā whakaahua; ehara i te kore-utu tētahi whakaahua-mata kiato, ā, ka kī tere ake te matapihi i tētahi wātea e whakaputa streaming ana i ngā papa i tētahi kuputuhi-anake noa. He tino tika te hoahanga mō tēnei: ko te tauira 5:1 matapihi-neke-ki-whānui te tikanga ka aro te nuinga o ngā paparanga i roto i te takiwā 512-tohu, ā, ka kukume mai anō ngā paparanga whānui i te hanganga pae-roa i ia wā.
07Te kaha whakaaroaro: utu mō te whakaaro i te wā anake e hiahiatia ana
Ka whakaatu a Inkling i te kaha whakaaroaro ka taea te whakahaere — tētahi o ngā pātene tino whai take o te rārangi. Ka whakaae te tauira o runga ake ki ngā taumata mai i te none me te minimal, mā te low, medium, high, xhigh, tae atu ki te max, e whakaaturia ana i ngā reo e toru mā te tawhā paerewa reasoning_effort.
{
"model": "inkling-W4A16-AUTOROUND-REAP",
"reasoning_effort": "low",
"messages": [ ... ]
}He rakaraka utu tino kaha tēnei, nā te mea ka utua ngā tohu whakaaroaro hei tohu whakaputa i te $6.07 ia miriona. E pūrongo ana a Thinking Machines i te Terminal Bench 2.1 ka eke a Inkling ki ngā tatauranga rite ki a Nemotron 3 Ultra i te tata hautoru o ngā tohu — he whāinga hoahanga i kīia te whaihua tohu mō te pūnaha kaha, ehara i te mea pōhēhē.
I roto i te mahi: whakahaerea te low, te minimal rānei mō te tapanga, te whakarōpū, me te tango kōrero māmā, i reira kei roto i te whakaahua te whakautu, ā, kāore he mekameka whakatau e hiahiatia ana. Rongoātia te high me ngā mea kei runga ake mō te whakaaroaro tūtohi, te whakataurite ataata maha-hikoi, me ngā hurihanga kaihautū i reira he utu nui tētahi hikoi-waenga hē. Karekau te nuinga o ngā tīma e whakarite i te tawhā, ā, ka waiho he rerekētanga utu-whakaputa whakareatanga-maha ki te tēpu.
08He aha te tikanga o "kore-ārai" i konei — ā, he aha kāore
Kia mārama: ka whakahaerea e mātou tēnei tauira me te kore paparanga whakakāhore, kore hoki tātari e pā ana ki te whakaputa. Kāore tētahi kaiwhakarōpū e pānui ana i tō tohutohu, e whakatau ana mēnā ka whakaaetia te tauira ki te whakautu, ā, kāore tētahi e pānui ana i te whakautu i mua i tana streaming. Ina rūrū, kore mōhio rānei te tauira, ka kī ia i tēnā — nā te tauira tēnā, ehara i te kaupapa here. Ko te mea kāore koe e whiwhi he whakakāhoretanga aunoa mō tētahi kaupapa e kīia ana e tētahi hua kaihoko-whānui he kaupapa here-tapu.
Mō te tirohanga, koirā te kaupapa uara katoa. He hapa te tika o ngā kaiwhakarōpū whakaahua i mua i ngā API tirohanga, ā, ka tau te reiti pono-teka ki te mahi ngaio e tino hiahia ana ki te āheinga: te rangahau haumarutanga, ngā horopaki haumanu, te kairīpoata me te taunakitanga, te arotake kiriata me te toi, ngā taputapu whakahaere e whai take ana ki te tirotiro i te mea e whakahaerea ana e ia. Kāore e taea e tētahi pito mutunga e whakakāhore ana i runga i te momo te whakahaere pono i tētahi o aua mea.
E rua ngā whakatūpato pono. Tuatahi, ka whakaputaina e Thinking Machines ngā arotake haumarutanga mō tā rātou whakatakotoranga tauira — 78.0% i te FORTRESS whakaraerae, 95.9% i te FORTRESS ngāwari, 98.6% i te StrongREJECT. Ka whakaatu ērā tau i tā rātou paparua me ō rātou ārai haumaru e mau tonu ana. Kāore rātou e whakaatu i tā mātou, ā, kāore mātou e whakahua i a rātou me te mea nā rātou. I tangohia e mātou te paparanga e inea wāhanga ana e aua tau; koirā te hua, koirā hoki te haepapa.
Tuarua, ehara te "kore paparanga whakakāhore" i te "kore ture." Ka whakahaerea a Shannon Lab LLC i raro i tētahi Kaupapa Here Whakamahi Haepapa me te ture US e pā ana. Ko te tango i tētahi kaiwhakarōpū he neke i te whakatau ki a koe, ehara i te muku i a ia. Mēnā kei te hanga koe i runga i tēnei, pānuitia te kaupapa here i mua i tō whakanui.
09Te tīmatanga
URL tūāpapa ōrite, kī ōrite, e toru ngā reo. Mēnā kei te karanga kē tō waehere i a OpenAI, Anthropic, i te Responses API rānei, ko te huri i te aho model te nuku katoa i te nuinga o te wā.
from openai import OpenAI
client = OpenAI(
base_url="https://api.shannon-ai.com/v1",
api_key=SHANNON_API_KEY,
)
stream = client.chat.completions.create(
model="inkling-W4A16-AUTOROUND-REAP",
messages=[{"role": "user", "content": [
{"type": "text", "text": "What does this figure actually show?"},
{"type": "image_url", "image_url": {"url": image_data_url}},
]}],
reasoning_effort="high",
max_tokens=4000,
stream=True,
)
for chunk in stream:
delta = chunk.choices[0].delta.content
if delta:
print(delta, end="")Kei te tuhinga API te tohutoro tawhā katoa, ngā āhua o /v1/messages me /v1/responses, ngā hōputu kaupae streaming, me tētahi papa whakamātau ora. Kei te Shannon Research ngā kāri tauira mō ērā atu tauira tekau mā tahi.
Ko ia tau tūāpapa me tikanga o runga ake nō tētahi pūtake tuatahi tūmatanui. Kāore he paearu i inea e Shannon i konei, ā, kāore mātou e whakaatu i tēnā — tirohia ngā mea taketake.
He tau i whakaputaina e Thinking Machines Lab mō te pauna Inkling kāore i kōpakina ngā tau tūāpapa, ā, e whakaatu ana i te tuanui o te whānau, kaua i tā mātou hanga quantized. He uara rārangi ngā utu me ngā tohu āheinga a te Kūwaha Shannon AI, e rite ana ki Hepetema 2026.
10Ngā pātai auau
He aha te API tauira Inkling kore-ārai?
Ko te whakatakotoranga a Shannon AI mō Inkling, te MoE taumaha-tuwhera 975B-tawhā i whakaputaina e Thinking Machines Lab i te 15 o Hurae 2026. Ka whakahaerea e mātou hei inkling-W4A16-AUTOROUND-REAP me te horopaki 262,144-tohu, te tāuru whakaahua, te karanga taputapu, me te whakahaere kaha whakaaroaro, i runga i tā mātou ake tautau GPU, me te kore paparanga whakakāhore, kore hoki tātari whakaputa.
Ka tautoko a Inkling i te tāuru whakaahua i runga i te API Shannon?
Āe — ko ia tētahi o ngā tauira e whā e whai-tirohanga ana i te rārangi. Tukua he wāhanga ihirangi image_url i runga i /v1/chat/completions, he poraka image rānei i runga i /v1/messages. Me te kore paparanga whakakāhore, ka whakaahuatia te mea kei roto i te papa, kaua e whakakāhoretia i runga i te momo.
Ka tautoko a Inkling i te putanga hanganga json_schema?
Kāo. Ka mahi te response_format mō te aratau ahanoa-JSON; kāore te json_schema kua whakaūngia-ā-wetereo e mahi. Whakamahia a Kimi-K2.6, a MiMo-V2.5 rānei mēnā e hiahiatia ana, ā, i runga i a Inkling whakamahia tētahi whakamārama taputapu akiaki hei schema māu, ka whakaū koe ake.
He aha te tikanga o W4A16-AUTOROUND-REAP?
Ngā taumaha 4-pī me ngā hohenga 16-pī; te AutoRound (Intel) e ako ana i te porowhita me te kutikuti mā te heke-paiaka-tohu kaua i te porowhita-ki-te-tata-rawa; me te REAP (Cerebras) te tapahi hikoi-kotahi i ngā mātanga MoE tatauranga-iti rawa kua whakatauhia ki ngā uara kēti kaiwhakarite ara me ngā ine whakaputa.
E hia te utu, ā, he whai take te utu nui ake?
$1.42 tāuru / $6.07 whakaputa ia miriona tohu — tata 3x tā mātou taumata papa, tata hoki te hautoru o te tauira tino teitei a Kimi-K3. He whai take mō te whakaaroaro tūtohi, ngā kaihautū whakaahua-mata, me ngā tuhinga kaha-whakaahua. Kāore he take mō te kuputuhi noa, te tango schema pukoro rānei.
He aha te matapihi horopaki?
262,144 tohu tāuru me te 16,384 tohu whakaputa ia tono, ōrite puta noa i ngā tauira tekau mā rua. Ka tuhia e te tūāpapa o runga ake te 1,048,576 tohu; ka whakahaerea e mātou he wāhi 262K kia kaua rawa te whanonga horopaki e noho hei take motuhake ā-tauira.
Whakamātauria te API Inkling kore-ārai
Tirohanga, horopaki 262K, ngā reo API e toru, kore paparanga whakakāhore.
Pānuitia ngā Tuhinga API Tirotiro i ngā Kāri Tauira KatoaShannon Lab LLC · New Mexico, USA · e whakahaerea ana i raro i tā mātou Kaupapa Here Whakamahi Haepapa
Ngā puna kōrero: Thinking Machines Lab — Te Whakauru mai o Inkling · Hugging Face — Nau mai a Inkling · REAP the Experts (arXiv 2510.13999) · Optimize Weight Rounding via Signed Gradient Descent (arXiv 2309.05516) · Sebastian Raschka — ngā tuhipoka hoahanga me paearu Inkling. He tau nā te whare rangahau i pūrongotia mō te pauna kāore i kōpakina ngā paearu tūāpapa.