API yeLaguna Isina Kudzorwa
Laguna S 2.1 — hwaro hweMixture-of-Experts hwe118B hunoshandiswa pasina chidzitiro chekuramba, hune hwindo recontext re262K, nekudana zvishandiso — pamutengo wepasi we$0.50 anopinda / $2.00 anobuda pamatokeni miriyoni.
Pfupiso
API yeLaguna isina kudzorwa inomhanyisa Laguna S 2.1 yePoolside — mapharamita 118B ose, anenge 8B anoshanda patokeni imwe neimwe, huremu hwakavhurika pasi peOpenMDW-1.1 — yakaitwa quantization kuenda kuW4A16 neAutoRound neREAP uye ichishandiswa pacluster yedu yemaGPU. Unowana context yematokeni 262,144, matokeni 16,384 anobuda pachikumbiro chimwe nechimwe, kuyerera, kudana zvishandiso, reasoning_effort, uye hapana sefa pane zvinobuda. Chinhu chimwe chinoyambirwa, chinotaurwa pakutanga: iyi ndiyo modhi yega mumutsara wedu isingatsigiri json_schema kana response_format. Kana uchida JSON yakavimbiswa, tora hanzvadzi — chivakwa chePro chaMiMo-V2.5 chinozviita pamutengo mumwe chetewo.
Mapeji mazhinji e"modhi yakachipa" anokukumbira kugamuchira hwaro hudiki, hwaneta, uchipiwa nhamba yakadzikira. Iri harizvidaro. Laguna S 2.1 ndiyo Mixture-of-Experts yakapararira yechizvarwa chazvino, yakadzidziswa pamaH200 4,096 uye yakaburitswa mumavhiki asingasviki mapfumbamwe, uye iri pamutengo wepasi we$0.50 / $2.00 mumwe chete nemodhi yega yega yemhando yekutanga yatinoshandisa. Kuchinjana kwakamanikana uye kuri nyore kufunga nezvako: unorasikirwa nezvinobuda zvakarongwa zvinomanikidzwa nesevha uchichengeta zvimwe zvose.
01Laguna S 2.1 chii?
Laguna S 2.1 imodhi ine huremu hwakavhurika yakaburitswa nePoolside musi wa21 Chikunguru 2026 pasi perezinesi rakasununguka OpenMDW-1.1 — nhengo yepakati yemhuri yeLaguna, iri pakati peLaguna XS 2.1 (33B/3B anoshanda) neLaguna M.1 (225B/23B anoshanda).
Magadzirirwo akabudiswa ndeerouter inosarudzwa netokeni ine gating yesoftplus panyanzvi 256 dzinotumirwa pamwe nenyanzvi imwe inogovaniswa, gumi dzichisarudzwa patokeni imwe neimwe, pamadanho 48 — gumi nemaviri eattention yakazara, makumi matatu nematanhatu ane hwindo rinotsvedza rematokeni 512, kurukanidzwa kunochengeta mutengo weattention wezvinopinda zvakareba zvikuru uri pedyo nekutwasuka pane kupeta kaviri. Attention iyeboka remibvunzo ine misoro yekiyi-nekukosha misere pachiyero chemusoro che128.
Zvinhu zviviri zvinobva ipapo zvinokosha kumushandisi weAPI. Kushanda kwakapararira ndicho chikonzero modhi ye118B ichigona kugara pamutengo wepasi — mapharamita anenge 8B chete ndiwo anoshanda patokeni, saka kuverenga patokeni kuri pedyo nemodhi diki yakazara pane zvinoratidzwa nehuwandu hwemapharamita. Uye kurukanidzwa kwehwindo rinotsvedza ndicho chikonzero hwindo refu richinyatsoshandiswa pane kungova pazita. Kuburitswa kwaPoolside kunotsigira matokeni 1,048,576; isu tinopa 262,144, zvakaenzana nemodhi yega yega paGedhi.
02Chinhu chimwe chinoyambirwa: hapana zvinobuda zvakarongwa
Laguna S 2.1 paShannon AI haitsigiri json_schema kana response_format. Ndiyo modhi yega mumutsara wedu wemamodhi gumi nemaviri isina kutsigira zvinobuda zvakarongwa; kutumira pharamita ipi zvayo yacho hakuzokupi mhinduro yakaganhurirwa neschema. Kudana zvishandiso kunoshanda. Kuyerera kunoshanda. Kufungisisa kunoshanda. Chinoshaikwa ndechekurongedza zvinobuda kunomanikidzwa nesevha chete.
Chichiri kushanda. Kudana zvishandiso kunotsigirwa zvizere, uye zvinopinda zvekudana chishandiso zvichiri kusvika sechinhu cheJSON zvakafanana nepamodhi ipi zvayo iri pano. Kana "zvinobuda zvakarongwa" zvechishandiso chako zviri chaizvo zvinopinda zvezvishandiso — zvinoita zvakadaro pamafuremuweki mazhinji emaagent, masevha eMCP nezvishandiso zvepamberi zvedhatabhesi — Laguna S 2.1 inokushandira zvakanaka. Kuyerera kunoshanda pamitauro mitatu yose; reasoning_effort inoshanda.
Chisingashandi. Haugoni kupa suwo JSON Schema uchivimbiswa kuti mhinduro yemavara yemubatsiri ichaienderana, uye response_format: {"type": "json_object"} haikutengeri chinhu. Kukumbira JSON kunowanzoburitsa JSON — iyi imodhi ine unyanzvi — asi "kunowanzo" hachisi chibvumirano. Poolside pachayo inonyora kunyorwa kusiriko kweJSON mukudana zvishandiso kwakadzama sechinhu chinozivikanwa chisingakwanisiki, chiyeuchidzo chekuti chivimbiso, apo chiripo pane dzimwe dziri mumutsara wedu, chinobva padanho rekushandisa pane pakufunga kwakanaka kwemodhi.
Hanzvadzi yekushandisa panzvimbo payo
Kana basa rako richida JSON yakavimbiswa, usarwe nemodhi iyi — chinja zita remodhi. Hanzvadzi ina dziri pamutengo mumwe chete we$0.50 / $2.00 dzine kutsigira kwakazara kwejson_schema neresponse_format:
| Zita remodhi | json_schema | Kuona | Sei waizoisarudza |
|---|---|---|---|
| MiMo-V2.5-Pro-W8A16 | Hongu | Kwete | Kuchinjana kwakananga — huremu hwe8-bit, mhando ine kutendeseka kwepamusoro, mutengo mumwe chete |
| MiMo-V2.5-W8A16 | Hongu | Hongu | Zvimwe chetezvo, pamwe nemifananidzo inopinda |
| Hy3-W8A16 | Hongu | Kwete | Imwe modhi ye8-bit yemavara pamutengo wepasi |
| DeepSeek-V4-Flash-0731 | Hongu | Kwete | W4A16-AutoRound-REAP seLaguna, kutsigira schema kuchiripo |
Sarudzo yechisikigo ndiMiMo-V2.5-Pro (W8A16): yemavara chete, iri pamutengo wepasi, yakaitwa quantization pamhando yakapfava kupfuura dzose dzatinoshandisa — tsananguro dziri pagwaro remodhi reMiMo-V2.5. Nokuti mamodhi gumi nemaviri ose anogovana masuwo, hwindo recontext nedenga rezvinobuda, kuchinjana ishanduko yemutsara mumwe pamunda wemodel.
03Chii chinorehwa neW4A16-AUTOROUND-REAP?
Chipfupiso chiri pazita remodhi hachisi chishongo — chinotaura chaizvo kuti huremu hwakadzvanywa sei husati hwaiswa pacluster yedu yemaGPU. Zvikamu zvitatu.
W4A16 — huremu hwe4-bit, maactivation e16-bit
Huremu hunochengetwa pamabhiti mana; maactivation — zvikoshwa zvinofamba pakati pemadanho panguva yekushandisa — anosara pa16. Kudimbura huremu kuita chikamu chechina chenzvimbo yahwo yeBF16 ndicho chinoita kuti modhi ye118B ikwane uye irambe ichimhanya, apo kuchengeta maactivation akafara kuchidzivirira chikamu chinonzwisisa nhamba chekupfuura kwemberi. Dzvanya chihombe uye chisingachinji, siya chidiki uye chinochinja.
AutoRound — kutenderedza kwakadzidzwa, kwete kutenderedza kune kuri pedyo
Nzira yakapusa ye quantization ndeyekutenderedza kune kuri pedyo: isa huremu humwe nehumwe pakukosha kwe4-bit kuri pedyo, gamuchira chikanganiso. AutoRound, seti yezvishandiso yakavhurika yeIntel ye quantization yemushure mekudzidzisa, inoita zviri nani — nzira yayo yeSignRound inoshandisa signed gradient descent kugadzirisa pamwe chete sarudzo dzekutenderedza nemiganhu yekucheka pamatanho anenge mazana maviri, ichisarudza pahuremu humwe nehumwe kuti kutenderedza mudenga kana pasi ndoku kunokuvadza zvishoma zvinobuda chaizvo zvedanho. Inotora pfungwa kubva pakudzidziswa kunoziva quantization pasina mutengo wekudzidziswazve, uye ndicho chikonzero chivakwa cheW4A16 chichigona kuva modhi chaiyo pane chinhu chakarasikirwa chinoshamisa.
REAP — kubvisa nyanzvi panzvimbo pekudzisanganisa
REAP iRouter-weighted Expert Activation Pruning, yakabudiswa neCerebras Research. Mumodhi yeMoE hadzisi nyanzvi dzose dzinoita basa: router inotumira matokeni mashoma kune dzimwe, uye dzimwe hadzibatsiri zvakawanda kunyange padzinoshanda. REAP inopa nyanzvi imwe neimwe chiyero pazviratidzo zvose zviri zviviri — zvikoshwa zvegedhi rerouter nezviyero zvekushanda kwenyanzvi — ipapo yobvisa dzisingakoshi kamwe chete, quantization isati yaitwa. Chinokosha ndechekuti inobvisa pane kusanganisa: kusanganisa kunopeta nyanzvi dzakawanda kuita imwe, kuchiparadza kudzora kwakatsetseka, kunotsamira pazvinopinda, kwerouter uye kuchiunza pasi pechikanganiso pasina kugadziriswa kunochibvisa (vanyori vanochidana kuti kuputsika kwenzvimbo yekushanda), apo kubvisa kuchisiya dzasara nekudzora kwerouter padziri zvakachengetedzeka. Cerebras inoshuma kubvisa kusvika ku50% yenyanzvi kubva pamamodhi emapharamita etiriyoni ichichengeta mhando yepakutanga — pamodhi yekukodha ye480B pakubvisa kwe50%, 97.6% yeunyanzvi hwekukodha husiri hwemaagent ne96.7% paSWE-Bench yemaagent. Idzo inhamba dzeCerebras dzemamodhi avo, kwete kuyera kwechivakwa chedu, asi ndicho chikonzero tichimhanyisa chimiro ichi pahwaro hune simba guru rekukodha.
Pamwe chete chipfupiso chinoverengwa semutsara: bvisa nyanzvi dzanga dzichingoshandiswa zvishoma, dzidza kutenderedza kwe4-bit pane kukufungidzira, chengeta maactivation akafara zvizere.
04Modhi iyi yakanakira chii chaizvo?
Chokwadi kutanga: Laguna S 2.1 imodhi yekutaurirana yemavara chete — hapana mifananidzo inopinda pazita remodhi iri — uye Poolside yakaivakira kukodha kwemaagent nebasa renguva refu, kwete mavara ekunyora. Kudzidziswa kwemashure kwakaitwa panzvimbo dzinopfuura 409,000, kusanganisira mabasa eterminal 83,000 neeunjiniya hwesoftware 168,000, aine reinforcement learning muFP8. Mhedzisiro yakashumwa naPoolside yekuburitswa kusina quantization:
| Muyero | Kufunga | Kusafunga | Chaunoyera |
|---|---|---|---|
| SWE-Bench Multilingual | 78.5% | — | Kugadzirisa zvikanganiso chaizvo mumitauro yakawanda |
| Terminal-Bench 2.1 | 70.2% | 60.4% | Basa reshell nezvishandiso rematanho akawanda |
| SWE-Bench Pro | 59.4% | — | Mabasa emaneja akaoma, asingasvibiswi |
| Toolathlon Verified | 49.7% | — | Kufara kwekushandisa zvishandiso pamaAPI akawanda |
| SWE Atlas | 46.2% | — | Kufamba mumaneja kwenguva refu |
| DeepSWE v1.1 | 40.4% | 16.5% | Unjiniya hwesoftware hwemaagent |
Mukaha uri pakati pemakadhi maviri ndiyo nhamba inobatsira kupfuura dzose pano: paDeepSWE chiyero chinopfuura kupetwa kaviri kana kufungisisa kwabvumirwa, uye paTerminal-Bench chinowedzera anenge mapoinzi gumi.
Nhamba dzakashumwa naPoolside dzekuburitswa kweLaguna S 2.1 kusina quantization, Chikunguru 2026. Hadzisi kuyera kwechivakwa chedu cheW4A16-AutoRound-REAP, uye hatiburitsi nhamba dzemiyero dzedu pachiri. Shandisa reasoning_effort kudzora bhajeti rekufunga paAPI yedu.
Pabasa rekutaurirana nerekusika, chinzvimbo chechokwadi ndechekuti: unyanzvi huri kumashure kwekumhanya kurefu kwemaagent — kubata chimiro chakawanda zvichinzwisisika, kuramba pabasa, kusatsauka — ndihwo humwe chetehwo huri kumashure kwebasa rekunyora hunhu hwenguva refu, uye pakuita inobata izwi zvakanaka pahwindo refu chaizvo. Hapana muyero wakabudiswa unodaro, uye hatizogadziri wedu. Iedze pamaprompt ako; pa$0.50 pamatokeni miriyoni anopinda izvozvo zvinenge zvemahara.
05Mutengo wepasi unotenga chii chaizvo?
$0.50 anopinda / $2.00 anobuda pamatokeni miriyoni ndiyo mhando yekutanga paGedhi reShannon AI, inogovaniswa neMiniMax-M3, DeepSeek-V4-Flash, MiMo-V2.5, MiMo-V2.5-Pro neHy3. Zviri pachena:
| Basa | Anopinda | Anobuda | Mutengo |
|---|---|---|---|
| Kutenderera kumwe kwehurukuro kune nhoroondo ye100K | 100,000 | 2,000 | $0.054 |
| Chitsauko chemazwi 5,000 kubva pagwaro refu | 30,000 | 7,000 | $0.029 |
| Kutenderera kwehurukuro kupfupi 1,000 | 4,000,000 | 500,000 | $3.00 |
| Kuverengwa kwecontext yakazara ye262K | 262,144 | 16,384 | $0.164 |
Mutsara wekupedzisira ndiwo waunofanira kunzwisisa: kuzadza hwindo rose recontext nekugadzira mhinduro yakanyanya yematokeni 16,384 kunodhura anenge masendi gumi nematanhatu. Kupa manyoro ose, zvakanyorwa zvegungano, kana seti yemagwaro mukudana kumwe chete kunorega kuva sarudzo yebhajeti kuchiita sarudzo yekunonoka. Masvomhu mamwe chetewo anotungamira rimwe basa remutengo wepasi — huwandu: kurongedza kweboka, kunyorazve kwakawanda, hurukuro dzekunyepedzera, kunyora kwepakutanga kunozogadziridzwa nemodhi yakasimba gare gare. Ona peji remitengo neindekisi yetsvakurudzo.
06Kuti "isina kudzorwa" inorevei pano
Zvinoreva zvazvinotaura: hapana chidzitiro chekuramba uye hapana sefa yezvirimo pane zvinobuda. Hapana muongorori ari pakati pemodhi nemhinduro yako achisarudza kana wakabvumirwa kubvunza; modhi inopindura mumutauro wawakumbira uye haikudzidzisi pamusoro pehwaro hwacho kutanga.
Pabasa rekusika ndiwo musiyano uri pakati pechishandiso chinoshanda nechisingashandi. Ngano dzine vapikisi vanotyisa, vanhu vanonyepa nemazvo, chisimba chine mhedzisiro, vanotaura nyaya vane tsika dzakakanganiswa — modhi inosiya hunhu ichiwedzera chinyorwa chekuzvidzivirira yakundikana pabasa, haina kubudirira pakuchengetedza. Zvimwe chetezvo zvinoshanda patsvakurudzo yezvekuchengetedza, zvinyorwa zvemapoka anopikisa, kunyora kwezvekurapa nezvepamutemo, nekushandura apo mavara ekutanga akajeka: zvikamu apo kuramba kunongova mhinduro isiriyo. Chazvisingarevi ndechekuti kutonga hakuchisiri kwako — unoramba uine mutoro wezvaunogadzira, pasi peMutemo wedu Wekushandisa Nemutsigo nemutemo unoshanda.
07Kudana API yeLaguna isina kudzorwa
Mitauro mitatu, modhi imwe chete: /v1/chat/completions (OpenAI), /v1/messages (Anthropic), /v1/responses. Kuyerera kunoshanda pamitatu yose, saka kubatanidzwa kuripo kunotamira nekungochinja base URL nezita remodhi — hapana kuchinja SDK, hapana kunyorazve kubata kwako meseji.
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"
}
Zvinyorwa zvishoma zvinoshanda:
- Usatumire
response_formatkanajson_schemakuzita remodhi iri — kana raibhurari yemutengi wako ichiisa imwe nechisikigo, ibvise kana uende kuhanzvadzi inoitsigira. - Zvishandiso zvinoenda muhurongwa hwe
toolshwakajairika — kudana zvishandiso kunotsigirwa zvizere; kumanikidzwa kweschema pamavara akasununguka chete ndiko kusipo. - Ronga bhajeti rako rekufungisisa.
reasoning_effortinoteererwa — yakadzikira kana yakadzimwa inochengeta kunonoka kwehurukuro kwakadzika nematokeni anobuda akachipa; ikwidze pabasa rezvishandiso rematanho akawanda, apo mukaha wemiyero uri pamusoro unoratidza zvaunotenga. - Denga rezvinobuda imatokeni 16,384 pachikumbiro. Pakugadzira kwakareba sebhuku, chikamurei muzvitsauko uzviyerese.
- Mavara chete. Hapana mifananidzo inopinda pazita remodhi iri — shandisa MiMo-V2.5, Kimi-K2.6, Kimi-K3, MiniMax-M3 kana inkling kana uchida kuona.
Rondedzero yakazara yemapharamita, kupinda, zvinorehwa nezvikanganiso nenzvimbo yekuedza iri kushanda: magwaro eAPI.
08Rini wekuisarudza, uye rini usingaisarudzi
Isarudze kana zvinobuda zvako zviri mavara ekunyora, hurukuro kana kodhi inoverengwa nemunhu; uchida kudana zvishandiso asi usingadi mavara akaganhurirwa neschema; mutengo patokeni uchitonga pane context refu; kana uchida pfungwa yechipiri kubva kudzinza risiri remhuri dzeDeepSeek, Kimi neMiMo dzinozadza chikamu chichiri kusara chemutsara.
Sarudza chimwe kana uchida JSON yakavimbiswa (MiMo-V2.5-Pro), kunzwisisa mifananidzo (Kimi-K2.6 kana MiMo-V2.5), kana mucheto wepamberi wekuoma apo mukaha wemutengo usina basa pamberi pekuwana mhinduro yakarurama (DeepSeek-V4-Pro, Kimi-K3).
json_schema neresponse_format. Verenga gwaro remodhi reMiMo-V2.5, kana tarisa gumi nemaviri ose muindekisi yetsvakurudzo.
Nharo dzose dzemagadzirirwo nedzemiyero dziri pano dzinobva kumuburitsi wemodhi kana kuvanyori venzira dzekudzvanya. Dzitarise iwe pachako.
09Mibvunzo inowanzobvunzwa
Chii chinonzi API yeLaguna isina kudzorwa?
NdiLaguna S 2.1 yePoolside ine huremu hwakavhurika — mapharamita mabhiriyoni zana negumi nemasere, anenge mabhiriyoni masere anoshanda patokeni imwe neimwe — ichishandiswa pacluster yeShannon yemaGPU pasina chidzitiro chekuramba uye pasina sefa yezvirimo pane zvinobuda. Zita remodhi ndiLaguna-S-2.1-W4A16-AUTOROUND-REAP, hwindo recontext imatokeni 262,144, uye inodhura $0.50 pamatokeni miriyoni anopinda ne$2.00 pamatokeni miriyoni anobuda.
Laguna S 2.1 inotsigira json_schema kana response_format paShannon AI here?
Kwete. Ndiyo modhi yega mumutsara weShannon isingagamuchiri chero imwe yacho. Kudana zvishandiso nekuyerera zvinoshanda zvakajairika uye zvinopinda zvekudana chishandiso zvichiri kusvika seJSON, asi hapana schema inomanikidzwa nesevha pamavara anobuda emubatsiri. Kana uchida JSON yakavimbiswa pamutengo mumwe chete we$0.50 / $2.00, shandisa MiMo-V2.5-Pro (W8A16).
Chii chinorehwa neW4A16-AUTOROUND-REAP?
W4A16 zvinoreva kuti huremu hunochengetwa pamabhiti mana apo maactivation achisara pamabhiti gumi nematanhatu. AutoRound inzira yeIntel ye quantization yemushure mekudzidzisa, inoshandisa signed gradient descent kudzidza sarudzo dzekutenderedza nedzekucheka panzvimbo pekutenderedza kune kuri pedyo. REAP ndiyo Router-weighted Expert Activation Pruning yeCerebras Research, inobvisa nyanzvi dzisingawanzotumirwi uye dzinotumirwa zvisina simba kubva pamodhi yeMixture-of-Experts quantization isati yaitwa pane kudzisanganisa.
Laguna S 2.1 yakanakira kunyora kwekusika nehurukuro here?
Imodhi yekutaurirana yemavara chete — hapana mifananidzo inopinda — uye inobata mavara marefu, anonzwisisika, ari muhunhu pahwindo re262K pasina chidzitiro chekuramba chiri munzira. Zvakanyorwa zvemiyero zvakabudiswa naPoolside zviri pakukodha kwemaagent, kwete pamavara ekunyora, saka edza mhando yekusika pamaprompt ako pane kuvimba nerondedzero.
API yeLaguna isina kudzorwa inodhura marii?
$0.50 pamatokeni miriyoni anopinda ne$2.00 pamatokeni miriyoni anobuda — mutengo wepasi paGedhi reShannon AI, unogovaniswa neMiniMax-M3, DeepSeek-V4-Flash, MiMo-V2.5, MiMo-V2.5-Pro neHy3. Hurukuro yematokeni 100,000 inogadzira matokeni 2,000 emhinduro inodhura anenge masendi 5.4.
Edza API yeLaguna isina kudzorwa
MoE ye118B · context ye262K · zvishandiso nekuyerera · $0.50 / $2.00 pamatokeni 1M.
Verenga Magwaro eAPI Tanga KutauriranaHapana chidzitiro chekuramba · hapana sefa pane zvinobuda · kutonga kwako, mutoro wako
Zvakatorwa kubva: Poolside — kadhi remodhi reLaguna S 2.1 · Introducing Laguna S 2.1 · REAP the Experts (Cerebras Research) · Intel AutoRound. Nhamba dzemagadzirirwo nedzemiyero dzakashumwa nemuburitsi dzekuburitswa kusina quantization, kwete kuyera kwechivakwa chatinoshandisa. Shannon AI inoshandiswa neShannon Lab LLC, New Mexico, USA.