I-API ye-Laguna Engacinezelwanga

Uninzi lwamaphepha "emodeli etshiphu" akucela ukuba wamkele isiseko esincinci, esidiniweyo ngenani elisezantsi. Eli alenzi njalo. I-Laguna S 2.1 kukukhutshwa kwe-Mixture-of-Experts esasazekileyo kwesi sizukulwana, eyaqeqeshwa kwii-H200 eziyi-4,096 yaza yahanjiswa ngaphantsi kweeveki ezilithoba, kwaye ihlala kumgangatho onye osezantsi we-$0.50 / $2.00 njengayo yonke enye imodeli yenqanaba lokungena esiyihambisayo. Intengiselwano imxinwa kwaye ilula ukuqiqa ngayo: uyalahla imveliso ecwangcisiweyo enyanzeliswa yiseva uze ugcine ezinye zonke.

01Yintoni i-Laguna S 2.1?

I-Laguna S 2.1 yimodeli enobunzima obuvulekileyo eyakhutshwa yi-Poolside nge-21 Julayi 2026 phantsi kwelayisensi evumelayo ye-OpenMDW-1.1 — ilungu eliphakathi losapho lwe-Laguna, phakathi kwe-Laguna XS 2.1 (33B/3B ezisebenzayo) ne-Laguna M.1 (225B/23B ezisebenzayo).

Isakhiwo esipapashiweyo yirawuta yokukhetha itoken nokuvula kwe-softplus phezu kweengcali ezingama-256 ezirawutwayo kunye nengcali enye eyabelwanayo, ezilishumi zikhethwa kwitoken nganye, kuzo zonke iimaleko ezingama-48 — ezilishumi elinambini zengqwalasela yehlabathi, ezingamashumi amathathu anesithandathu zefestile etyibilikayo kwifestile yeetoken ezingama-512, uphithikezo olugcina ixabiso lengqwalasela lezinto ezingenayo ezinde kakhulu likufuphi ekutheni tye endaweni yokuphindwa kabini. Ingqwalasela ngeyombuzo oqoqiweyo neentloko zesitshixo-nexabiso ezisibhozo kubungakanani bentloko obuyi-128.

Iziphumo ezimbini zibalulekile kumthengi we-API. Ukusebenza okusasazekileyo kusisizathu sokuba imodeli ye-118B ikwazi ukuhlala kwixabiso elisezantsi — iiparamitha ezimalunga nezingama-8B kuphela ezisebenzayo kwitoken nganye, ngoko ukubala kwitoken nganye kukufuphi kwimodeli encinci eshinyeneyo kunokuba inani leeparamitha libonisa. Kwaye uphithikezo lwefestile etyibilikayo lusisizathu sokuba ifestile ende isebenziseke ngokwenene endaweni yokuba ibe yeyamagama. Ukukhutshwa kwe-Poolside kuxhasa iitoken ezingama-1,048,576; thina sihambisa ezingama-262,144, kuhambelana nayo yonke enye imodeli kwiSango.

02Isilumkiso esinye: akukho mveliso icwangcisiweyo

Into esasebenzayo. Ukubiza izixhobo kuxhaswe ngokupheleleyo, kwaye iingxoxo zokubiza izixhobo zisafika njengento ye-JSON kanye njengakuyo nayiphi na enye imodeli apha. Ukuba "imveliso ecwangcisiweyo" yesicelo sakho ngokwenene ziingxoxo zezixhobo — into eyinyani kuninzi lweenkqubo zeemeli, iiseva ze-MCP nezingaphambili zoovimba bedatha — i-Laguna S 2.1 iyakusebenzela kakuhle. Ukusasaza kuyasebenza kuzo zontathu iilwimi; i-reasoning_effort iyasebenza.

Into engasebenziyo. Awunako ukunika i-endpoint isikimu se-JSON uze uqinisekiswe ukuba impendulo yombhalo yomncedisi iyaqinisekiseka ngokuchasene naso, kwaye i-response_format: {"type": "json_object"} ayikuthengeli nto. Ukuyalela i-JSON kudla ngokuvelisa i-JSON — le yimodeli enobuchule — kodwa "ukudla ngoku" ayisosivumelwano. I-Poolside ngokwayo ibhala ukubaleka okungachanekanga kwe-JSON kwizibizo zezixhobo ezinzulu njengomda owaziwayo, isikhumbuzo sokuba isiqinisekiso, apho sikhona kwenye indawo kuluhlu lwethu, sivela kumaleko wokuhambisa endaweni yeenjongo ezilungileyo zemodeli.

Nguwuphi udade omawusetyenziswe endaweni yayo

Ukuba umsebenzi wakho ufuna i-JSON eqinisekisiweyo, musa ukulwa nale modeli — tshintsha isazisi semodeli. Oodade abane bahlala kwixabiso elifanayo le-$0.50 / $2.00 benenkxaso epheleleyo ye-json_schema neye-response_format:

Isazisi semodeli	json_schema	Umbono	Kutheni ubuya kuyikhetha
MiMo-V2.5-Pro-W8A16	Ewe	Hayi	Utshintshiselwano oluthe ngqo — ubunzima beebhithi ezisibhozo, inqanaba elithembeke kunazo zonke, ixabiso elinye
MiMo-V2.5-W8A16	Ewe	Ewe	Into enye, kunye nomfanekiso ongenayo
Hy3-W8A16	Ewe	Hayi	Enye imodeli yombhalo yeebhithi ezisibhozo kwixabiso elisezantsi
DeepSeek-V4-Flash-0731	Ewe	Hayi	I-W4A16-AutoRound-REAP njengeLaguna, inkxaso yesikimu isaphelele

Ukhetho olungagqibekanga lwemvelo yi-MiMo-V2.5-Pro (W8A16): ngumbhalo kuphela, inexabiso elisezantsi, icuthwe kwinqanaba elithambileyo kunazo zonke esizisebenzisayo — iinkcukacha zikwi-sikhokelo seModeli ye-MiMo-V2.5. Kuba zonke iimodeli ezilishumi elinambini zabelana ngee-endpoint, ifestile yomxholo nophahla lwemveliso, ukutshintsha lutshintsho lomgca omnye kwindawo ye-model.

03Ithetha ntoni i-W4A16-AUTOROUND-REAP?

Impumezelo kwisazisi semodeli ayilohombiso — ichaza kanye indlela ubunzima obacinezelwa ngayo phambi kokuba bulayishwe kwiqoqo lethu leeGPU. Zintathu iinxalenye.

W4A16 — ubunzima beebhithi ezine, ukusebenza kweebhithi ezilishumi elinesithandathu

Ubunzima bugcinwa kwiibhithi ezine; ukusebenza — amaxabiso ahamba phakathi kweemaleko ngexesha lokuqikelela — buhlala kwezilishumi elinesithandathu. Ukunciphisa ubunzima buye kwisiqingatha sesine sobungakanani baso be-BF16 kokwenza imodeli ye-118B ilingane ihlale ikhawuleza, ngoxa ukugcina ukusebenza kubanzi kukhusela inxalenye ebuthathaka ngamanani yodlulo olungaphambili. Cinezela into enkulu nemileyo, ushiye into encinci neguqukayo.

AutoRound — ukujikeleza okufundiweyo, hayi ukujikelezela kweyona ikufuphi

Indlela engenabuchule yokucutha kukujikelezela kweyona ikufuphi: bopha ubunzima ngabunye kwelona xabiso lifuphi leebhithi ezine, wamkele impazamo. I-AutoRound, ithuluzi elivulekileyo le-Intel lokucutha emva koqeqesho, yenza ngcono — indlela yayo ye-SignRound isebenzisa ukwehla kwesantya okunophawu ukulungiselela kunye izigqibo zokujikeleza neenqanaba zokunqumla kumanyathelo amakhulu amabini, igqiba ubunzima ngabunye ukuba ukujikelezela phezulu okanye phantsi kwenza umonakalo omncinci kwimveliso yokwenene yemaleko. Iboleka umbono kuqeqesho oluqonda ukucutha ngaphandle kwexabiso loqeqesho olutsha, kwaye sisizathu sokuba isakhiwo se-W4A16 sinokuba yimodeli enzulu endaweni yokuba yinto elahlekelweyo enomdla nje.

REAP — ukuthena iingcali endaweni yokuzidibanisa

I-REAP yi-Router-weighted Expert Activation Pruning, epapashwe yi-Cerebras Research. Kwimodeli ye-MoE ayizizo zonke iingcali ezitsala umthwalo wazo: irawuta ithumela iitoken ezimbalwa kwezinye, kwaye ezinye zifaka isandla esincinci ngobungakanani xa zidubula. I-REAP inika inqaku ingcali nganye kuzo zombini iimpawu — amaxabiso egeyithi yerawuta kunye nemilinganiselo yokusebenza kweengcali — ize isuse ezinokubaluleka okusezantsi kwisitshotsho esinye, phambi kokucutha. Okubalulekileyo iyathena endaweni yokudibanisa: ukudibanisa kusonga iingcali ezininzi zibe nye, kutshabalalise ulawulo olucokisekileyo, oluxhomekeke kwizinto ezingenayo lwerawuta kwaye kwazise umgangatho wempazamo ongasuswa kukutuna (ababhali bakubiza ngokuwa kwendawo yokusebenza), ngoxa ukuthena kushiya ezisindileyo nolawulo lwerawuta phezu kwazo lusaphelele. I-Cerebras ixela ukususa ukuya kuma-50% eengcali kwiimodeli zeeparamitha ezisithriliyoni ngoxa kugcinwa umgangatho osisiseko — kwimodeli yokukhowuda ye-480B kukuthena kwe-50%, ama-97.6% esakhono sokukhowuda esingesiso se-agentic esisisiseko nama-96.7% kwi-SWE-Bench ye-agentic. Lawo ngamanani e-Cerebras kwiimodeli zayo, ingekuko amalinganiselo esakhiwo sethu, kodwa sisizathu sokuba siqhube esi sitaki kwisiseko ekuqina kwaso okuphambili yikhowudi.

Kunye impumezelo ifundeka njengesivakalisi: thena iingcali ebezisetyenziswa kancinci, funda ukujikeleza kweebhithi ezine endaweni yokuqikelela, ugcine ukusebenza kububanzi obupheleleyo.

04Ilungele ntoni ngokwenene le modeli?

Ubunyaniso kuqala: i-Laguna S 2.1 yimodeli yencoko yombhalo kuphela — akukho mfanekiso ongenayo kwesi sazisi semodeli — kwaye i-Poolside yayakhela ukukhowuda kwe-agentic nomsebenzi wexesha elide, hayi umbhalo. Emva koqeqesho kwaqhutywa kwiindawo ezingama-409,000, kubandakanya imisebenzi yetheminali engama-83,000 neyobunjineli besoftwe engama-168,000, ngokufunda okuqinisekisiweyo kwi-FP8. Iziphumo ezixelwe yi-Poolside ngokukhutshwa okungacutshwanga:

Ubalo-mgangatho	Ukucinga	Ngaphandle kokucinga	Ilinganisa ntoni
SWE-Bench Multilingual	78.5%	—	Izilungiso zeempazamo zokwenene kwiilwimi ezininzi
Terminal-Bench 2.1	70.2%	60.4%	Umsebenzi weqokobhe nezixhobo wamanyathelo amaninzi
SWE-Bench Pro	59.4%	—	Imisebenzi yovimba enzima ngakumbi, echasa ukungcoliseka
Toolathlon Verified	49.7%	—	Ububanzi bokusetyenziswa kwezixhobo kwii-API ezininzi
SWE Atlas	46.2%	—	Ukuhamba kovimba kwexesha elide
DeepSWE v1.1	40.4%	16.5%	Ubunjineli besoftwe be-agentic

Umsantsa phakathi kwamakholomu amabini lelona nani lisebenzayo apha: kwi-DeepSWE inqaku liphinda liphindwe ngaphezu kabini xa ukuqiqa kuvuliwe, kwaye kwi-Terminal-Bench lizuza amanqaku aphantse abe lishumi.

Amanani axelwe yi-Poolside ngokukhutshwa kwe-Laguna S 2.1 okungacutshwanga, Julayi 2026. Ayingomalinganiselo esakhiwo sethu se-W4A16-AutoRound-REAP, kwaye asipapashi manani obalo-mgangatho ethu ngaso. Sebenzisa i-reasoning_effort ukulawula uhlahlo-lwabiwo-mali lokucinga kwi-API yethu.

Kumsebenzi wencoko nowokuyila isimo esinyanisekileyo sesi: iimpawu ezisemva kobaleko olude lwe-agentic — ukubamba ubume obuninzi ngokuhambelanayo, ukuhlala kumsebenzi, ukungaphambuki — zezinye ezisemva komsebenzi omde womlinganiswa, kwaye ngokwenza ibamba ilizwi kakuhle kwifestile ende kakhulu. Akukho balo-mgangatho lupapashiweyo olutshoyo, kwaye asizukuyila. Yivavanye kwimiyalelo yakho; nge-$0.50 kwitoken enye yezigidi ezingenayo oko kuphantse kube simahla.

05Lithenga ntoni ngokwenene ixabiso elisezantsi?

I-$0.50 engenayo / $2.00 ephumayo kwitoken enye yezigidi linqanaba lokungena kwiSango le-Shannon AI, elabelwana ne-MiniMax-M3, i-DeepSeek-V4-Flash, i-MiMo-V2.5, i-MiMo-V2.5-Pro ne-Hy3. Ngokuthe ngqo:

Umsebenzi	Ezingenayo	Eziphumayo	Ixabiso
Umjikelo omnye wencoko onembali yeetoken ezingama-100K	100,000	2,000	$0.054
Isahluko samagama angama-5,000 kwingxelo emfutshane ende	30,000	7,000	$0.029
Imijikelo yencoko emifutshane eliwaka	4,000,000	500,000	$3.00
Ukufundwa komxholo opheleleyo weetoken ezingama-262K	262,144	16,384	$0.164

Umqolo wokugqibela ngowokumbelwa ngaphakathi: ukuzalisa yonke ifestile yomxholo nokuvelisa impendulo ephezulu yeetoken ezingama-16,384 kubiza malunga neesenti ezilishumi elinesithandathu. Ukondla umbhalo opheleleyo, umbhalo weseshoni, okanye iseti yamaxwebhu kumbizo omnye kuyeka ukuba sisigqibo sohlahlo-lwabiwo-mali size sibe sisigqibo sexesha lokulinda. Izibalo ezinye ziqhuba esinye isicelo sexabiso elisezantsi — umthamo: ukwahlula ngeendidi ngeqela, ukubhala kwakhona ngobuninzi, incoko yokwenziwa, ukubhalwa kokuqala imodeli enamandla eyicokisayo kamva. Bona iphepha lamaxabiso kunye nalfabhethi yophando.

06Kuthetha ntoni "ukungacinezelwa" apha

Kuthetha oko kukutshoyo: akukho maleko wokwala kwaye akukho kuhluzwa kwesiqulatho kwimveliso. Akukho sahluli-ndidi sihleli phakathi kwemodeli nempendulo yakho sigqiba ukuba ubuvunyelwe na ukubuza; imodeli iphendula ngesimo osicelileyo yaye ayikufundisi ngesiseko kuqala.

Kumsebenzi wokuyila loo ngumahluko phakathi kwesixhobo esisebenzayo nesingasebenziyo. Intsomi inabachasi abasongelayo, abalinganiswa abaxoka ngokukholekayo, ubundlobongela obunemiphumo, ababalisi abonakeleyo ngokwesimilo — imodeli eyaphula umlinganiswa ukuze yongeze isilumkiso isilele kumsebenzi, ayiphumelelanga kukhuseleko. Kunjalo nakuphando lokhuseleko, amaxwebhu eqela elibomvu, ukubhalwa kwezonyango nezomthetho, kunye nokuguqulelwa apho umthombo ungqalileyo: iindidi apho ukwala kuyimpendulo engachanekanga nje. Into engakuthethiyo kukuba isigwebo siyeka ukuba sesakho — usanoxanduva ngento oyiveliswayo, phantsi kwePolisi yethu yoSetyenziso oluNoxanduva nomthetho osebenzayo.

07Indlela yokubiza i-API ye-Laguna engacinezelwanga

Iilwimi ezintathu, imodeli enye: i-/v1/chat/completions (OpenAI), i-/v1/messages (Anthropic), i-/v1/responses. Ukusasaza kuyasebenza kuzo zontathu, ngoko udibaniso olukhoyo ludlulela apha ngokutshintsha i-URL esisiseko nesazisi semodeli — akukho tshintshiselwano lwe-SDK, akukho kubhalwa kwakhona kokuphathwa kwemiyalezo yakho.

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"
}

Amanqaku ambalwa asebenzayo:

• Musa ukuthumela i-response_format okanye i-json_schema kwesi sazisi semodeli — ukuba ithala lomthengi wakho limisela enye ngokungagqibekanga, yisuse okanye uye kudade wayo oyixhasayo.
• Izixhobo zihamba kuluhlu oluqhelekileyo lwe-tools — ukubiza izixhobo kuxhaswe ngokupheleleyo; kuphela ukunyanzeliswa kwesikimu kumbhalo okhululekileyo okungekhoyo.
• Yenza uhlahlo-lwabiwo-mali lokuqiqa kwakho. I-reasoning_effort iyahlonitshwa — okusezantsi okanye okuvaliweyo kugcina ixesha lokulinda lencoko lisezantsi neetoken eziphumayo zitshiphu; yinyuse ngomsebenzi wezixhobo wamanyathelo amaninzi, apho umsantsa wobalo-mgangatho ongasentla ubonisa into oyithengayo.
• Uphahla lwemveliso ziitoken ezingama-16,384 kwisicelo ngasinye. Ngokuvelisa ubungakanani bencwadi, yenza izahluko uze usasaze.
• Ngumbhalo kuphela. Akukho mfanekiso ongenayo kwesi sazisi semodeli — sebenzisa i-MiMo-V2.5, i-Kimi-K2.6, i-Kimi-K3, i-MiniMax-M3 okanye i-inkling ngombono.

Ireferensi epheleleyo yeeparamitha, ukuqinisekiswa, intsingiselo yeempazamo nendawo yokudlala ephilayo: amaxwebhu e-API.

08Ixesha lokuyikhetha, nexesha lokungayikhethi

Yikhethe xa imveliso yakho ingumbhalo, incoko okanye ikhowudi efundwa ngumntu; udinga ukubiza izixhobo kodwa hayi umbhalo othintelwe sisikimu; ixabiso letoken nganye lilawula phezu komxholo omde; okanye ufuna uluvo lwesibini kumnombo ongengowe-DeepSeek, i-Kimi nezosapho lwe-MiMo ezizalisa okuseleyo koluhlu.

Khetha enye into xa ufuna i-JSON eqinisekisiweyo (MiMo-V2.5-Pro), ukuqonda umfanekiso (Kimi-K2.6 okanye MiMo-V2.5), okanye icala lomphambili lobunzima apho umsantsa wexabiso ungabalulekanga xa uthelekiswa nokufumana impendulo echanekileyo (DeepSeek-V4-Pro, Kimi-K3).

09Imibuzo ebuzwa rhoqo

---

Imithombo: Poolside — isikhokelo seModeli ye-Laguna S 2.1 · Introducing Laguna S 2.1 · REAP the Experts (Cerebras Research) · Intel AutoRound. Amanani esakhiwo nawobalo-mgangatho axelwe ngumpapashi ngokukhutshwa okungacutshwanga, ingekuko amalinganiselo esakhiwo esisihambisayo. I-Shannon AI iqhutywa yi-Shannon Lab LLC, New Mexico, USA.