I-API ye-Laguna Engavinjelwe

Amakhasi amaningi e-"modeli eshibhile" akucela ukuthi wamukele isisekelo esincane, esikhathele ngokushintshaniswa nenani eliphansi. Leli akunjalo. I-Laguna S 2.1 ingukukhishwa kwesizukulwane samanje kwe-Mixture-of-Experts esabalele, okwaqeqeshwa kuma-H200 angu-4,096 futhi kwathunyelwa ngaphansi kwamasonto ayisishiyagalolunye, futhi ihlala ephansini elifanayo lika-$0.50 / $2.00 njengayo yonke enye imodeli yezinga lokungena esiyikhonzayo. Ukushintshana kuncane futhi kulula ukucabanga ngakho: udela okukhishwayo okuhlelekile okuphoqwa yiseva bese ugcina konke okunye.

01Iyini i-Laguna S 2.1?

I-Laguna S 2.1 iyimodeli yezisindo ezivulekile eyakhishwa yi-Poolside ngo-Julayi 21, 2026 ngaphansi kwelayisensi evumelanayo ye-OpenMDW-1.1 — ilungu elimaphakathi lomndeni we-Laguna, phakathi kwe-Laguna XS 2.1 (33B/3B esebenzayo) ne-Laguna M.1 (225B/23B esebenzayo).

Ukwakhiwa okushicilelwe yi-router ye-token-choice ene-softplus gating phezu kochwepheshe abangu-256 abaqondiswayo kanye nochwepheshe oyedwa owabelwanayo, abayishumi bakhethwa kuthokheni ngalinye, kumangqimba angu-48 — ayishumi nambili okunaka komhlaba wonke, angamashumi amathathu nesithupha e-sliding-window ewindini lamathokheni angu-512, ukuxutshwa okugcina izindleko zokunaka kokufakwayo okude kakhulu kuseduze nokuqonda kunokuphindwa kabili. Ukunaka kungukwe-grouped-query ngezinhloko zekhiye-nanani eziyisishiyagalombili ngobukhulu benhloko obungu-128.

Imiphumela emibili ibalulekile kumsebenzisi we-API. Ukusebenza okusabalele yikho okwenza imodeli engu-118B ikwazi ukuhlala entengweni ephansi kunazo zonke — amapharamitha acishe abe ngu-8B kuphela asebenza kuthokheni ngalinye, ngakho ukubala kuthokheni ngalinye kuseduze nemodeli encane eminyene kunalokho inani lamapharamitha elikushoyo. Futhi ukuxutshwa kwe-sliding-window yikho okwenza iwindi elide lisebenziseke ngempela kunokuba libe ligama nje. Ukukhishwa kwe-Poolside kusekela amathokheni angu-1,048,576; thina sikhonza angu-262,144, kufana nayo yonke enye imodeli esangweni.

02Isixwayiso esisodwa: akukho okukhishwayo okuhlelekile

Okusasebenza. Ukubiza amathuluzi kusekelwa ngokugcwele, futhi izimpikiswano zokubiza amathuluzi zisafika njengento ye-JSON ncamashi njengakunoma yiyiphi enye imodeli lapha. Uma u-"okukhishwayo okuhlelekile" kohlelo lwakho empeleni kuyizimpikiswano zamathuluzi — okuyiqiniso ezinhlakeni eziningi zama-ejenti, kumaseva e-MCP nakuzo zonke izinhlelo zangaphambili zesizindalwazi — i-Laguna S 2.1 iyakusiza kahle. Ukusakaza kuyasebenza kuzo zontathu izilimi; u-reasoning_effort uyasebenza.

Okungasebenzi. Awukwazi ukunikeza isiphetho i-JSON Schema bese uqinisekiswa ukuthi impendulo yembhalo yomsizi iyaqinisekiswa ngokumelene nayo, futhi u-response_format: {"type": "json_object"} awukuthengeli lutho. Ukucela i-JSON ngokuvamile kukhiqiza i-JSON — le yimodeli enekhono — kodwa u-"ngokuvamile" akusiyo isivumelwano. I-Poolside uqobo ibhala ukubalekiswa okungalungile kwe-JSON ekubizweni kwamathuluzi okujule kakhulu njengomkhawulo owaziwayo, isikhumbuzo sokuthi isiqinisekiso, lapho sikhona kwenye indawo ohlwini lwethu, sivela ongqimbeni lokukhonza kunasezinhlosweni ezinhle zemodeli.

Yimuphi umfowethu ongasebenzisa esikhundleni salokho

Uma umsebenzi wakho udinga i-JSON eqinisekisiwe, ungalwi nale modeli — shintsha i-id yemodeli. Abafowethu abane bahlala entengweni efanayo ka-$0.50 / $2.00 ngokusekelwa okugcwele kwe-json_schema nokwe-response_format:

I-id yemodeli	json_schema	Umbono	Kungani ungayikhetha
MiMo-V2.5-Pro-W8A16	Yebo	Cha	Ukushintshana okuqondile — izisindo zamabhithi ayisishiyagalombili, izinga elethembeke kunawo wonke, intengo efanayo
MiMo-V2.5-W8A16	Yebo	Yebo	Okufanayo, kanye nokufakwa kwesithombe
Hy3-W8A16	Yebo	Cha	Enye imodeli yombhalo yamabhithi ayisishiyagalombili ngentengo ephansi kunazo zonke
DeepSeek-V4-Flash-0731	Yebo	Cha	W4A16-AutoRound-REAP njenge-Laguna, ukusekelwa kwesikimu kusekhona

Okuzenzakalelayo okwemvelo yi-MiMo-V2.5-Pro (W8A16): ingeyombhalo kuphela, inentengo ephansi kunazo zonke, ikaliwe ezingeni elithambe kunawo wonke esiliqhubayo — imininingwane isekhadini lemodeli le-MiMo-V2.5. Ngenxa yokuthi wonke amamodeli ayishumi nambili abelana ngeziphetho, ngewindi lomongo nangophahla lokukhishwayo, ukushintsha kungushintsho lomugqa owodwa enkanjini ye-model.

03Kusho ukuthini u-W4A16-AUTOROUND-REAP?

Isijobelelo se-id yemodeli asiwona umhlobiso — sisho ncamashi ukuthi izisindo zacindezelwa kanjani ngaphambi kokulayishwa eqoqweni lethu lama-GPU. Izingxenye ezintathu.

W4A16 — izisindo zamabhithi amane, ukusebenza kwamabhithi ayi-16

Izisindo zigcinwa emabhithini amane; ukusebenza — amanani ageleza phakathi kwamangqimba ngesikhathi sokuqagela — kuhlala kwayi-16. Ukunciphisa izisindo zibe yingxenye eyodwa kwezine yesikhala sazo se-BF16 yikho okwenza imodeli engu-118B ilingane futhi ihlale isheshe, kanti ukugcina ukusebenza kubanzi kuvikela ingxenye ezwela ngezinombolo yokudlula okuya phambili. Cindezela okukhulu nokungaguquki, ushiye okuncane nokuguqukayo kungathintwanga.

AutoRound — ukugoba okufundiwe, hhayi ukugoba okuseduze

Indlela elula yokukala yi-round-to-nearest: bopha isisindo ngasinye enanini lamabhithi amane eliseduze, wamukele iphutha. I-AutoRound, ithuluzi elivulekile le-Intel lokukala ngemuva kokuqeqeshwa, yenza kangcono — indlela yayo ye-SignRound isebenzisa i-signed gradient descent ukulungiselela ndawonye izinqumo zokugoba namabanga okunquma ezinyathelweni ezingamakhulu ambalwa, inquma esisindweni ngasinye ukuthi ukugoba phezulu noma phansi kwenza umonakalo omncane kokukhishwayo kwangempela kwengqimba. Ithatha umqondo ekuqeqeshweni okuqaphela ukukalwa ngaphandle kwezindleko zokuqeqeshwa kabusha, futhi yingakho ukwakhiwa kwe-W4A16 kungaba yimodeli engathi esiqotho kunokuba yinto elahlekelwe.

REAP — ukuthena ochwepheshe esikhundleni sokubahlanganisa

I-REAP yi-Router-weighted Expert Activation Pruning, eyashicilelwa yi-Cerebras Research. Kwimodeli ye-MoE akubona bonke ochwepheshe abadonsa isisindo sabo: i-router ithumela amathokheni ambalwa kwabanye, futhi abanye bafaka ubukhulu obuncane uma besebenza. I-REAP inikeza uchwepheshe ngamunye amaphuzu kuzo zombili izimpawu — amanani egeti ye-router kanye namazinga okusebenza kochwepheshe — bese isusa abanokubaluleka okuphansi ngokushaya kanye, ngaphambi kokukala. Okubalulekile ukuthi iyathena kunokuhlanganisa: ukuhlanganisa kugoqa ochwepheshe abaningana babe munye, kubhubhise ukulawula okuncane, okuxhomeke kokufakwayo kwe-router futhi kwethule iphansi lephutha okungekho ukushuna okulisusayo (ababhali bakubiza ngokuthi ukuwa kwesikhala sokusebenza), kanti ukuthena kushiya abasalayo nokulawula kwe-router phezu kwabo kungathintekile. I-Cerebras ibika ukususa kufika ku-50% wochwepheshe emamodelini amapharamitha ayisigidintathu kuyilapho igcina ikakhulu ikhwalithi eyisisekelo — kwimodeli yekhodi engu-480B ekuthenweni kwama-50%, u-97.6% wekhono lekhodi eliyisisekelo elingelona ele-ejenti no-96.7% ku-SWE-Bench ye-ejenti. Lawo ngamanani e-Cerebras emamodelini abo, hhayi ukukalwa kokwakhiwa kwethu, kodwa yiwona sizathu sokuthi siqhube lesi sitaki phezu kwesisekelo amandla aso ayinhloko okuyikhodi.

Ndawonye isijobelelo sifundeka njengomusho: thena ochwepheshe abebesetshenziswa kancane, ufunde ukugoba kwamabhithi amane kunokukuqagela, ugcine ukusebenza kubanzi ngokugcwele.

04Yini le modeli eyilungele ngempela?

Iqiniso kuqala: i-Laguna S 2.1 iyimodeli yengxoxo yombhalo kuphela — akukho okufakwayo okuyisithombe kule id yemodeli — futhi i-Poolside yayakhela ikhodi ye-ejenti nomsebenzi wesikhathi eside, hhayi umbhalo oqambekile. Ukuqeqeshwa kwangemuva kwaqhutshwa ezindaweni ezingaphezu kuka-409,000, kufaka imisebenzi engu-83,000 ye-terminal nengu-168,000 yobunjiniyela besofthiwe, ngokufunda okuqinisayo ku-FP8. Imiphumela ebikwe yi-Poolside ngokukhishwa okungakaliwe:

Ukuhlolwa	Ngokucabanga	Ngaphandle kokucabanga	Kukala ini
SWE-Bench Multilingual	78.5%	—	Ukulungiswa kwamabhugi angempela ezilimini eziningi
Terminal-Bench 2.1	70.2%	60.4%	Umsebenzi we-shell namathuluzi onezinyathelo eziningi
SWE-Bench Pro	59.4%	—	Imisebenzi enzima kakhulu yendawo yokugcina ikhodi emelana nokungcoliswa
Toolathlon Verified	49.7%	—	Ububanzi bokusetshenziswa kwamathuluzi kuma-API amaningi
SWE Atlas	46.2%	—	Ukuzulazula endaweni yokugcina ikhodi isikhathi eside
DeepSWE v1.1	40.4%	16.5%	Ubunjiniyela besofthiwe be-ejenti

Umehluko phakathi kwezinhlu ezimbili uyinombolo esebenzayo kunazo zonke lapha: ku-DeepSWE iphuzu liphindeka ngaphezu kwakabili uma ukucabanga kuvuliwe, kanti ku-Terminal-Bench lizuza cishe amaphuzu ayishumi.

Amanani abikwe yi-Poolside ngokukhishwa kwe-Laguna S 2.1 okungakaliwe, Julayi 2026. Akusikho ukukalwa kokwakhiwa kwethu kwe-W4A16-AutoRound-REAP, futhi asishicileli manani okuhlolwa awethu ngakho. Sebenzisa u-reasoning_effort ukulawula isabelo sokucabanga ku-API yethu.

Emsebenzini wengxoxo nowokuqamba isimo esiqotho yilesi: izimfanelo ezingemuva kokusebenza okude kwe-ejenti — ukubamba isimo esiningi ngokuhambisanayo, ukuhlala emsebenzini, ukungaphambuki — yizona zimfanelo ezingemuva komsebenzi wesikhathi eside wobuntu bomlingiswa, futhi empeleni ibamba izwi kahle ewindini elide kakhulu. Akukho ukuhlolwa okushicilelwe okusho lokho, futhi ngeke sikuqambe. Kuhlole emiyalelweni yakho; ku-$0.50 ngethokheni eyisigidi engenayo lokho cishe kumahhala.

05Ithengani ngempela intengo ephansi kunazo zonke?

U-$0.50 engenayo / $2.00 ephumayo ngethokheni eyisigidi uyizinga lokungena esangweni le-Shannon AI, elabelwana ne-MiniMax-M3, i-DeepSeek-V4-Flash, i-MiMo-V2.5, i-MiMo-V2.5-Pro ne-Hy3. Ngokuqondile:

Umsebenzi	Okungenayo	Okuphumayo	Izindleko
Umzuliswano owodwa wengxoxo onomlando wamathokheni angu-100K	100,000	2,000	$0.054
Isahluko samagama angu-5,000 kusuka kwisifinyezo eside	30,000	7,000	$0.029
Imizuliswano emifushane yengxoxo engu-1,000	4,000,000	500,000	$3.00
Ukufundwa komongo ogcwele wamathokheni angu-262K	262,144	16,384	$0.164

Umugqa wokugcina yiwona okufanele uwugcine ekhanda: ukugcwalisa lonke iwindi lomongo nokukhiqiza impendulo ephezulu yamathokheni angu-16,384 kubiza cishe amasenti angu-16. Ukufaka umbhalo wonke, ukubhalwe phansi kweseshini, noma isethi yamadokhumenti ekubizeni okukodwa kuyeka ukuba yisinqumo sesabelomali kube yisinqumo sesikhathi sokuphendula. Izibalo ezifanayo ziqhuba enye indawo yokusebenzisa yentengo ephansi kunazo zonke — ivolumu: ukuhlukanisa ngeqoqo, ukubhala kabusha ngobuningi, ingxoxo yokwenziwa, nokubhala isilinganiso sokuqala imodeli enamandla kakhulu ezosilungisa kamuva. Bheka ikhasi lamanani kanye nenkomba yocwaningo.

06Kusho ukuthini u-"ongavinjelwe" lapha

Kusho lokho okushiwoyo: alukho ungqimba lokwenqaba futhi akukho ukuhlungwa kokuqukethwe kokukhishwayo. Asikho isihlukanisi esihlala phakathi kwemodeli nempendulo yakho sinquma ukuthi wawuvunyelwe yini ukubuza; imodeli iphendula ngesitayela osicelile futhi ayikufundisi ngesizathu kuqala.

Emsebenzini oqambekile lowo ngumehluko phakathi kwethuluzi elisebenzayo nelingasebenzi. Izinganekwane ziqukethe abaphikisi abesabekayo, abalingiswa abaqamba amanga ngokukholisayo, udlame olunemiphumela, abaxoxi abanezimilo ezingcolile — imodeli ephula ubuntu bomlingiswa ukuze ifake isaziso yehlulekile umsebenzi, ayiphumelelanga ekuphepheni. Okufanayo kubambeka kucwaningo lokuphepha, kumibhalo yeqembu elibomvu, ekubhaleni kwezokwelapha nokwezomthetho, nasekuhumusheni lapho umthombo uqonde ngqo: izigaba lapho ukwenqaba kuvele kube yimpendulo engalungile. Okungasho kona ukuthi ukwahlulela kuyayeka ukuba ngokwakho — uhlala unesibopho ngalokho okukhiqizayo, ngaphansi kwe-Nqubomgomo Yethu Yokusebenzisa Ngokuzibophezela nomthetho osebenzayo.

07Ungayibiza kanjani i-API ye-Laguna engavinjelwe

Izilimi ezintathu, imodeli eyodwa: /v1/chat/completions (OpenAI), /v1/messages (Anthropic), /v1/responses. Ukusakaza kuyasebenza kuzo zontathu, ngakho ukuhlanganiswa okukhona kudlulela lapha ngokushintsha i-URL eyisisekelo ne-id yemodeli — akukho ukushintshwa kwe-SDK, akukho ukubhalwa kabusha kokuphathwa kwemilayezo 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"
}

Amanothi ambalwa asebenzayo:

• Ungathumeli u-response_format noma u-json_schema kule id yemodeli — uma umtapo weklayenti lakho umisa omunye ngokuzenzakalelayo, wususe noma udlulele kumfowethu owusekelayo.
• Amathuluzi ahamba ku-array ejwayelekile ye-tools — ukubiza amathuluzi kusekelwa ngokugcwele; ukuphoqwa kwesikimu kumbhalo okhululekile kuphela okungasekelwa.
• Hlelela ukucabanga kwakho. U-reasoning_effort uyahlonishwa — okuphansi noma okuvaliwe kugcina isikhathi sokuphendula sengxoxo siphansi namathokheni okukhishwayo eshibhile; wukhuphule emsebenzini wamathuluzi onezinyathelo eziningi, lapho umehluko wokuhlolwa ongenhla ukhombisa ukuthi uthengani.
• Uphahla lokukhishwayo lungamathokheni angu-16,384 esicelweni ngasinye. Ekukhiqizeni okubude bencwadi, hlukanisa ngezahluko bese usakaza.
• Umbhalo kuphela. Akukho okufakwayo okuyisithombe kule id yemodeli — sebenzisa i-MiMo-V2.5, i-Kimi-K2.6, i-Kimi-K3, i-MiniMax-M3 noma i-inkling ngombono.

Ireferensi ephelele yamapharamitha, ukuqinisekiswa, incazelo yamaphutha kanye nendawo yokudlala ephilayo: imibhalo ye-API.

08Nini ukuyikhetha, nokuthi nini ungayikhethi

Yikhethe uma okukhishwayo kwakho kungumbhalo, ingxoxo noma ikhodi efundwa ngumuntu; udinga ukubiza amathuluzi kodwa hhayi umbhalo oboshwe yisikimu; izindleko ngethokheni ngalinye zibusa emongweni omude; noma ufuna umbono wesibili wozalo olungeyilo olwemindeni ye-DeepSeek, i-Kimi ne-MiMo egcwalisa okusele ohlwini.

Khetha okunye uma udinga i-JSON eqinisekisiwe (MiMo-V2.5-Pro), ukuqonda izithombe (i-Kimi-K2.6 noma i-MiMo-V2.5), noma uhlangothi lomkhawulo lobunzima lapho umehluko wentengo ungabalulekile uma kuqhathaniswa nokuthola impendulo elungile (i-DeepSeek-V4-Pro, i-Kimi-K3).

09Imibuzo ebuzwa njalo

---

Imithombo: Poolside — ikhadi lemodeli le-Laguna S 2.1 · Introducing Laguna S 2.1 · REAP the Experts (Cerebras Research) · Intel AutoRound. Amanani okwakhiwa nawokuhlolwa abikwe ngumshicileli ngokukhishwa okungakaliwe, hhayi ukukalwa kokwakhiwa esikukhonzayo. I-Shannon AI iqhutshwa yi-Shannon Lab LLC, New Mexico, USA.