API Laguna Kore-ārai
Laguna S 2.1 — he tūāpapa Whakaranu-o-Ngā-Mātanga 118B e whakahaerea ana me te kore paparanga whakakāhore, he matapihi horopaki 262K, me te karanga taputapu — i te utu papa $0.50 tāuru / $2.00 whakaputa ia miriona tohu.
Whakarāpopoto
Ka whakahaere te API Laguna kore-ārai i a Laguna S 2.1 a Poolside — 118B tawhā katoa, ~8B hohe ia tohu, taumaha tuwhera i raro i te OpenMDW-1.1 — kua quantized ki te W4A16 me AutoRound me REAP, ā, e whakahaerea ana i runga i tā mātou ake tautau GPU. Ka riro mai i a koe he horopaki 262,144-tohu, 16,384 tohu whakaputa ia tono, te streaming, te karanga taputapu, te reasoning_effort, me te kore tātari whakaputa. Kotahi whakatūpato, kia mōhio wawe: koinei anake te tauira o tā mātou rārangi kāore e tautoko ana i te json_schema, i te response_format rānei. Mēnā he hiahia tōu ki te JSON kua whakaūngia, tangohia tētahi tuakana/teina — ka taea e te hanga Pro a MiMo-V2.5 tēnā i te utu ōrite.
Ko te nuinga o ngā whārangi "tauira utu-iti" e tono ana kia whakaae koe ki tētahi tūāpapa iti, ngenge, hei utu mō tētahi nama iti. Ehara tēnei i tēnā. He putanga Whakaranu-o-Ngā-Mātanga korekore o tēnei whakatupuranga a Laguna S 2.1, i whakangungua ki runga i te 4,096 H200, ā, i tukuna i roto i te iwa wiki, ā, e noho ana ia ki te papa $0.50 / $2.00 ōrite ki ia tauira taumata-tāuru atu e whakahaeretia ana e mātou. He whakawhitinga kuiti, māmā hoki te whakaaro: ka tukua atu e koe te putanga hanganga kua whakaūngia e te tūmau, ā, ka mau tonu ngā mea katoa atu.
01He aha a Laguna S 2.1?
He tauira taumaha-tuwhera a Laguna S 2.1, i whakaputaina e Poolside i te 21 o Hurae 2026 i raro i te raihana taritari OpenMDW-1.1 — te mema rahi-waenga o te whānau Laguna, i waenganui i a Laguna XS 2.1 (33B/3B hohe) me Laguna M.1 (225B/23B hohe).
Ko te hoahanga i whakaputaina he kaiwhakarite-ara kōwhiri-tohu me te kēti softplus i runga i te 256 mātanga kua whakaritea-ara me te kotahi mātanga tiritahi, tekau kua kōwhiria ia tohu, puta noa i te 48 paparanga — tekau mā rua aro-whānui, toru tekau mā ono aro matapihi-neke i tētahi matapihi 512-tohu, he whakawhitiwhiti e pupuri ana i te utu aro o ngā tāuru tino roa kia tata-rārangi kaua e wharite. He aro-pātai-kāhui te aro, me te 8 pane-KV i te rahinga pane 128.
E rua ngā hua e nui ana ki tētahi kaiwhakamahi API. Ko te whakaohonga korekore te take ka taea ai tētahi tauira 118B te noho ki te utu papa — ~8B tawhā anake e mahi ana ia tohu, nō reira he tata ake te tatauranga ia tohu ki tētahi tauira kiato iti i te mea e kīia ana e te tatauranga tawhā. Ā, ko te whakawhitiwhiti matapihi-neke te take he tino whai-take te matapihi roa, kaua he mea ingoa noa. Ka tautoko te putanga a Poolside i te 1,048,576 tohu; ka whakahaerea e mātou te 262,144, e rite ana ki ia tauira atu o te kūwaha.
02Te whakatūpato kotahi: kore putanga hanganga
Kāore a Laguna S 2.1 i runga i a Shannon AI e whakaae ana ki te json_schema, ki te response_format rānei. Koinei anake te tauira o tā mātou rārangi tekau mā rua kāore he tautoko putanga-hanganga; ki te tuku koe i tētahi o ēnei tawhā kāore koe e whiwhi i tētahi whakautu kua herea ki te schema. Ka mahi te karanga taputapu. Ka mahi te streaming. Ka mahi te whakaaroaro. Ko te whakaahua putanga kua whakaūngia e te tūmau anake te mea kāore.
Ko ngā mea e mahi tonu ana. Ka tino tautokona te karanga taputapu, ā, ka tae tonu mai ngā tohutohu karanga-taputapu hei ahanoa JSON, rite tonu ki tētahi atu tauira o konei. Mēnā ko te "putanga hanganga" o tō taupānga he tohutohu taputapu tūturu — he pono tēnei mō te nuinga o ngā anga kaihautū, ngā tūmau MCP me ngā mata-pae pātengi raraunga — ka tino pai a Laguna S 2.1 ki a koe. Ka mahi te streaming i ngā reo e toru; ka mahi te reasoning_effort.
Ko ngā mea kāore e mahi. Kāore e taea e koe te hoatu i tētahi JSON Schema ki te pito mutunga me te mautohu ka whakaūngia te whakautu-kuputuhi a te kaiāwhina ki reira, ā, kāore he painga o te response_format: {"type": "json_object"} ki a koe. Ki te tono koe mō te JSON, he JSON tāna e hanga ai i te nuinga o te wā — he tauira kaha tēnei — engari ehara te "i te nuinga o te wā" i te kirimana. E tuhi ana a Poolside anō i te mawhiti hē o te JSON i ngā karanga taputapu hōhonu rawa hei herenga kua mōhiotia, hei whakamaumahara nā te paparanga whakahaere, kaua nā ngā hiahia pai o te tauira, te mautohu e puta ana i ērā atu wāhi o tā mātou rārangi.
Ko tēhea tuakana/teina hei whakamahi
Mēnā e hiahia ana tō kawenga mahi ki te JSON kua whakaūngia, kaua e whawhai ki tēnei tauira — hurihia te id tauira. E whā ngā tuakana/teina kei te utu ōrite $0.50 / $2.00 me te tautoko katoa mō te json_schema me te response_format:
| Id tauira | json_schema | Tirohanga | He aha koe ka kōwhiri ai |
|---|---|---|---|
| MiMo-V2.5-Pro-W8A16 | Āe | Kāo | Te whakakapi tika — taumaha 8-pī, taumata pono-rawa, utu ōrite |
| MiMo-V2.5-W8A16 | Āe | Āe | Rite tonu, me te tāuru whakaahua |
| Hy3-W8A16 | Āe | Kāo | Tētahi atu tauira kuputuhi 8-pī i te utu papa |
| DeepSeek-V4-Flash-0731 | Āe | Kāo | He W4A16-AutoRound-REAP rite ki a Laguna, e mau tonu ana te tautoko schema |
Ko te taunoa māori ko MiMo-V2.5-Pro (W8A16): kuputuhi-anake, utu-papa, kua quantized ki te taumata ngāwari rawa e whakahaeretia ana e mātou — ngā taipitopito kei te kāri tauira MiMo-V2.5. Nā te mea ka tiritahi ngā tauira tekau mā rua katoa i ngā pito mutunga, te matapihi horopaki me te tuanui whakaputa, he panoni kotahi-rārangi noa ki te mata model te huri.
03He aha te tikanga o W4A16-AUTOROUND-REAP?
Ehara te taunga o te id tauira i te whakapaipai noa — e kī tika ana i te āhua i kōpakina ai ngā taumaha i mua i tō rātou utaina ki runga i tā mātou tautau GPU. E toru ngā wāhanga.
W4A16 — ngā taumaha 4-pī, ngā hohenga 16-pī
Ka rokohia ngā taumaha ki te 4 pī; ka noho tonu ngā hohenga — ngā uara e rere ana i waenga i ngā paparanga i te wā whakatau — ki te 16. Ko te tapahi i ngā taumaha ki te hauwhā o tō rātou horahanga BF16 te mea e uru ai, e tere tonu ai hoki te rere o tētahi tauira 118B, engari mā te whānui tonu o ngā hohenga e tiaki ana i te wāhanga tino whakaaro-tatau o te rere-whakamua. Kōpakina te mea nui, te mea kore-neke; waiho anake te mea iti, te mea hurihuri.
AutoRound — te porowhita kua akona, kaua te porowhita-ki-te-tata-rawa
Ko te ara ngāwari ki te quantize he porowhita-ki-te-tata-rawa: kohia ia taumaha ki te uara 4-pī tata rawa, whakaae ki te hapa. Ka pai ake a AutoRound, te rauemi quantization-whai-muri-whakangungu tuwhera a Intel — ka whakamahi tāna tikanga SignRound i te heke-paiaka-tohu hei whakatikatika ngātahi i ngā whakatau porowhita me ngā awhe kutikuti puta noa i te rua rau hikoi, e whakatau ana ia taumaha mēnā he iti ake te kino o te porowhita-ake, porowhita-iho rānei ki te whakaputa tūturu a te paparanga. Ka nama ia i te whakaaro mai i te whakangungu mataara-quantization me te kore utu o te whakangungu-anō, ā, koinā te take ka taea ai tētahi hanga W4A16 te noho hei tauira taimaha, kaua he mea ngoikore e ngaro ai te mōhiohio.
REAP — te tapahi mātanga kaua te whakakotahi
Ko REAP te Router-weighted Expert Activation Pruning, i whakaputaina e Cerebras Research. I roto i tētahi tauira MoE, kāore ngā mātanga katoa e whai take ōrite: he iti ngā tohu ka tukuna e te kaiwhakarite ara ki ētahi, ā, he iti te mahi a ētahi ina puta. Ka whakatau te REAP i ia mātanga i runga i ngā tohu e rua — ngā uara kēti kaiwhakarite ara me ngā ine whakaputa mātanga — ā, ka tango i ngā mea tatauranga-iti rawa i te hikoi kotahi, i mua i te quantization. Te mea nui, ka tapahi kaua ka whakakotahi: ka whakakotahi te whakakotahi i ngā mātanga maha ki te kotahi, e whakangaro ana i te mana whakahaere mārama, ā-tāuru a te kaiwhakarite ara, ā, ka whakauru i tētahi papa hapa kāore e tangohia e te whakatikatika (e kīia ana e ngā kaituhi he "horonga mokowā-mahi"), engari ka waiho e te tapahi ngā mea e ora tonu ana me te mana whakahaere a te kaiwhakarite ara ki a rātou. E pūrongo ana a Cerebras i te tango tae atu ki te 50% o ngā mātanga mai i ngā tauira taera-tawhā me te pupuri tonu i te nuinga o te kounga paerewa — i tētahi tauira waehere 480B i te 50% tapahi, 97.6% o te kaha waehere-kore-kaihautū paerewa me te 96.7% i te SWE-Bench kaihautū. He tau nā Cerebras ērā mō ā rātou ake tauira, ehara i te mea i inea e mātou, engari koirā te take ka whakahaerea ai e mātou tēnei tūāpapa e kaha rawa ana ki te waehere.
Ka pānui te taunga katoa hei rerenga kōrero kotahi: tapahi i ngā mātanga i tata kore whakamahia, ako i te porowhita 4-pī kaua e whakapae, waiho ngā hohenga ki te whānui katoa.
04He aha tino te mea e pai ai tēnei tauira?
Kia pono i te tuatahi: he tauira kōrerorero kuputuhi-anake a Laguna S 2.1 — kāore he tāuru whakaahua i tēnei id tauira — ā, i hangaia e Poolside mō te waehere kaihautū me te mahi pae-roa, kaua mō te tuhinga. I rere te whakangungu-whai-muri puta noa i te 409,000 taiao, tae atu ki te 83,000 mahi anga me te 168,000 mahi hangarau pūmanawa, me te RL i te FP8. Ngā hua i pūrongotia e Poolside mō te putanga kāore i kōpakina:
| Paearu | Whakaaro | Kore-whakaaro | He aha te mea e inea ana |
|---|---|---|---|
| SWE-Bench Multilingual | 78.5% | — | Ngā whakatikatika hapa tūturu puta noa i ngā reo maha |
| Terminal-Bench 2.1 | 70.2% | 60.4% | Mahi anga me te taputapu maha-hikoi |
| SWE-Bench Pro | 59.4% | — | Ngā mahi pātaka uaua ake, aukati-poke |
| Toolathlon Verified | 49.7% | — | Te whānui whakamahi taputapu puta noa i ngā API maha |
| SWE Atlas | 46.2% | — | Te whakatere pātaka pae-roa |
| DeepSWE v1.1 | 40.4% | 16.5% | Hangarau pūmanawa kaihautū |
Ko te āputa i waenga i ngā tīwae e rua te tau tino whai-take o konei: i te DeepSWE, neke atu i te whakareatanga-rua te tatauranga ina whakahohea te whakaaroaro, ā, i te Terminal-Bench ka piki tata tekau piro.
He tau i pūrongotia e Poolside mō te putanga Laguna S 2.1 kāore i kōpakina, Hurae 2026. Ehara ērā i te mea i inea e mātou i tā mātou hanga W4A16-AutoRound-REAP, ā, kāore mātou e whakaputa i ā mātou ake tau paearu mō taua hanga. Whakamahia te reasoning_effort hei whakahaere i te pūtea whakaaro i runga i tā mātou API.
Mō te mahi kōrerorero me te auaha, koinei te tūranga pono: ko ngā āhuatanga kei muri i ngā rere kaihautū roa — te pupuri i te tūnga nui me te kōherenga, te noho ki te mahi, te kore kōtiti — ko ērā tonu ngā āhuatanga kei muri i te mahi āhua-tangata whitinga-roa, ā, i roto i te mahi, ka pai tana pupuri i te reo puta noa i tētahi matapihi tino roa. Kāore he paearu i whakaputaina e kī ana i tēnā, ā, kāore mātou e hanga i tētahi. Whakamātauria ki runga i āu ake tohutohu; i te $0.50 ia miriona tohu tāuru, he tata kore-utu tēnā.
05He aha tino ka hokona e te utu papa?
Ko te $0.50 tāuru / $2.00 whakaputa ia miriona tohu te taumata tāuru i runga i te Kūwaha Shannon AI, e tiritahi ana me MiniMax-M3, DeepSeek-V4-Flash, MiMo-V2.5, MiMo-V2.5-Pro me Hy3. Kia tino mārama:
| Kawenga mahi | Tāuru | Whakaputa | Utu |
|---|---|---|---|
| Kotahi hurihanga kōrero me te 100K hītori | 100,000 | 2,000 | $0.054 |
| He upoko 5,000-kupu mai i tētahi tuhinga tohutohu roa | 30,000 | 7,000 | $0.029 |
| 1,000 hurihanga kōrerorero poto | 4,000,000 | 500,000 | $3.00 |
| He pānui horopaki 262K-tohu katoa | 262,144 | 16,384 | $0.164 |
Ko te rārangi whakamutunga te mea kia mau ki roto i a koe: ko te whakakī i te matapihi horopaki katoa me te hanga i te whakautu 16,384-tohu tuanui rawa ka tata tekau mā ono heneti te utu. Ko te whāngai i tētahi tuhinga-ā-ringa katoa, tuhinga-kōrero hui, huinga tuhinga rānei ki roto i te karanga kotahi, ka mutu te noho hei whakatau pūtea, ka noho hei whakatau roa-wā noa iho. Ko taua tatauranga tonu e akiaki ana i tētahi atu take mō te utu papa — te rahi: te whakarōpū pukoro, te tuhi anō pukoro, te kōrerorero waihanga, te tuhinga hukihuki tuatahi mā tētahi tauira kaha ake e whakatika ai i muri. Tirohia te whārangi utu me te tauwhāinga rangahau.
06He aha te tikanga o "kore-ārai" i konei
Ko te tikanga tonu tāna e kī ana: kore paparanga whakakāhore, kore hoki tātari ihirangi e pā ana ki te whakaputa. Kāore he kaiwhakarōpū e noho ana i waenga i te tauira me tō whakautu e whakatau ana mēnā i whakaaetia koe ki te pātai; ka whakautu te tauira i te reo i tonoa e koe, ā, kāore ia e kauwhau ki a koe mō te pūtake i mua.
Mō te mahi auaha, koinā te rerekētanga i waenga i tētahi taputapu whai-take me tētahi kore-whai-take. He whaiaro-whawhai kino, he āhua tangata e teka pono ana, he tūkino me ōna hua, he kaikōrero kua ngoikore te matatika kei roto i te tuhinga-parau — ko tētahi tauira e wāhi ana i tana āhua ki te tāpiri whakapāha kua hapa i te mahi, kāore i angitu ki te haumaru. He pērā anō mō te rangahau haumarutanga, te tuhinga tīma-whero, te tuhinga rongoā me te ture, me te whakamāori i te wā he māro te pūtake: he akomanga i reira ko te whakakāhoretanga he whakautu hē noa. Ehara te tikanga o tēnei i te mea kua mutu tō whakatau — ka noho tonu koe hei kaiwhakahaere mō tāu e hanga ana, i raro i tā mātou Kaupapa Here Whakamahi Haepapa me te ture e pā ana.
07Me pēhea te karanga i te API Laguna kore-ārai
E toru ngā reo, kotahi te tauira: /v1/chat/completions (OpenAI), /v1/messages (Anthropic), /v1/responses. Ka mahi te streaming i ngā mea e toru, nā reira ka neke tētahi whakaurunga o nāianei mā te huri i te URL tūāpapa me tētahi id tauira — kāore he whakakapi SDK, kāore he tuhi-anō i tō whakahaere karere.
POST https://api.shannon-ai.com/v1/chat/completions
{
"model": "Laguna-S-2.1-W4A16-AUTOROUND-REAP",
"messages": [
{ "role": "system", "content": "You are a hardboiled narrator. Stay in voice." },
{ "role": "user", "content": "Open chapter nine." }
],
"stream": true,
"max_tokens": 4096,
"reasoning_effort": "low"
}
He iti ngā tuhipoka whai take:
- Kaua e tuku i te response_format, i te json_schema rānei ki tēnei id tauira — mēnā ka whakaritea tētahi e tō rārangi pūkete taunoa, wetekina, neke rānei ki tētahi tuakana/teina e tautoko ana i tēnā.
- Ka uru ngā taputapu ki te rārangi tools pēnei tonu — ka tino tautokona te karanga taputapu; ko te whakaū schema mō te kuputuhi kore-here anake kāore.
- Whakaritea tō pūtea whakaaroaro. Ka whakamanahia te
reasoning_effort— mā te iti, te weto rānei e iti tonu ai te roa-wā kōrerorero me te utu-iti o ngā tohu whakaputa; whakaneketia mō te mahi taputapu maha-hikoi, i reira e whakaatu ana te āputa paearu o runga ake i te hua o tēnā. - Ko te tuanui whakaputa he 16,384 tohu ia tono. Mō te hanga rahi-pukapuka, wāhia hei upoko, kātahi ka streaming.
- Kuputuhi anake. Kāore he tāuru whakaahua i tēnei id tauira — whakamahia a MiMo-V2.5, Kimi-K2.6, Kimi-K3, MiniMax-M3, inkling rānei mō te tirohanga.
Kei te tuhinga API te tohutoro tawhā katoa, te whakamanatanga, te tikanga hapa, me tētahi papa whakamātau ora.
08Āhea te kōwhiri, āhea kaua
Kōwhiria ina he tuhinga, he kōrerorero, he waehere rānei tō whakaputa ka pānuitia e te tangata; e hiahia ana koe ki te karanga taputapu engari kaua ki te kuputuhi herea-ki-te-schema; ka nui ake te utu ia-tohu i te horopaki roa; he hiahia tōu ki tētahi whakaaro tuarua mai i tētahi whakapapa kē atu i ngā whānau DeepSeek, Kimi me MiMo e whakakī ana i te toenga o te rārangi.
Kōwhiria tētahi atu ina e hiahia ana koe ki te JSON kua whakaūngia (MiMo-V2.5-Pro), te mārama whakaahua (Kimi-K2.6, MiMo-V2.5 rānei), te pito uaua rawa o te mata i reira kāore he take te āputa utu i te taha o te tika o te whakautu (DeepSeek-V4-Pro, Kimi-K3).
Ko ia kī hoahanga me te paearu o konei nā te kaiwhakaputa o te tauira, nā ngā kaituhi o ngā tikanga kōpaki rānei. Tirotirohia e koe ake.
09Ngā pātai auau
He aha te API Laguna kore-ārai?
Ko te taumaha-tuwhera a Poolside, a Laguna S 2.1 — 118 piriona tawhā, tata ki te 8 piriona hohe ia tohu — e whakahaerea ana i runga i tā Shannon ake tautau GPU me te kore paparanga whakakāhore, kore hoki tātari whakaputa. Ko te id tauira ko Laguna-S-2.1-W4A16-AUTOROUND-REAP, ko te matapihi horopaki he 262,144 tohu, ā, he $0.50 ia miriona tohu tāuru me te $2.00 ia miriona tohu whakaputa te utu.
Ka tautoko a Laguna S 2.1 i te json_schema, i te response_format rānei i runga i a Shannon AI?
Kāo. Koinei anake te tauira o te rārangi Shannon kāore e whakaae ana ki tētahi o ēnei. Ka mahi paerewa te karanga taputapu me te streaming, ā, ka tae tonu mai ngā tohutohu karanga-taputapu hei JSON, engari kāore he schema kua whakaūngia e te tūmau mō te whakaputa kuputuhi a te kaiāwhina. Mō te JSON kua whakaūngia i te utu ōrite $0.50 / $2.00, whakamahia a MiMo-V2.5-Pro (W8A16).
He aha te tikanga o W4A16-AUTOROUND-REAP?
Ko te tikanga o W4A16 kua rokohanga ngā taumaha ki te 4 pī, engari ka noho tonu ngā hohenga ki te 16 pī. Ko AutoRound te tikanga quantization-whai-muri-whakangungu a Intel, e whakamahi ana i te heke-paiaka-tohu hei ako i ngā whakatau porowhita me te kutikuti, kaua i te porowhita-ki-te-tata-rawa. Ko REAP te Router-weighted Expert Activation Pruning a Cerebras Research, e tango ana i ngā mātanga onge me te ngoikore te whakaritea-ara mai i tētahi tauira Whakaranu-o-Ngā-Mātanga i mua i te quantization, kaua e whakakotahi ana i a rātou.
He pai a Laguna S 2.1 mō te tuhinga auaha me te kōrerorero?
He tauira kōrerorero kuputuhi-anake — kāore he tāuru whakaahua — ā, ka mau ia i tētahi kuputuhi roa, kōherea, ū-ki-te-āhua puta noa i tētahi matapihi 262K me te kore paparanga whakakāhore e ārai ana. Kei te waehere kaihautū te rekoata paearu i whakaputaina e Poolside, kaua ki te kuputuhi, nō reira me whakamātau i te kounga auaha ki runga i āu ake tohutohu, kaua e whakawhirinaki ki tētahi papa tohu.
E hia te utu o te API Laguna kore-ārai?
$0.50 ia miriona tohu tāuru me te $2.00 ia miriona tohu whakaputa — te utu papa i runga i te Kūwaha Shannon AI, e tiritahi ana me MiniMax-M3, DeepSeek-V4-Flash, MiMo-V2.5, MiMo-V2.5-Pro me Hy3. Ko tētahi kōrerorero 100,000-tohu e hua ana i te 2,000 tohu whakautu ka tata $0.054 te utu.
Whakamātauria te API Laguna kore-ārai
MoE 118B · horopaki 262K · taputapu me te streaming · $0.50 / $2.00 ia 1M tohu.
Pānuitia ngā Tuhinga API Tīmatahia te KōreroreroKore paparanga whakakāhore · kore tātari whakaputa · nāu te whakatau, nāu te haepapa
Ngā puna kōrero: Poolside — kāri tauira Laguna S 2.1 · Te whakauru mai o Laguna S 2.1 · REAP the Experts (Cerebras Research) · Intel AutoRound. He tau i pūrongotia e te kaiwhakaputa mō te putanga kāore i kōpakina ngā tau hoahanga me te paearu, ehara i te mea i inea mō te hanga e whakahaeretia ana e mātou. Ka whakahaerea a Shannon AI e Shannon Lab LLC, New Mexico, USA.