Laguna API Yosasidwa

Matsamba ambiri a "mtundu wotsika mtengo" amakufunsani kuti muvomereze maziko aang'ono, otopa posinthana ndi manambala otsika. Iyi si imodzi mwa izo. Laguna S 2.1 ndi kutulutsidwa kwa mbadwo wapano wa Mixture-of-Experts yosakanikirana, yophunzitsidwa pa ma H200 4,096 ndipo inatumizidwa pansi pa milungu isanu ndi inayi, ndipo ili pa mtengo wotsika kwambiri womwewo wa $0.50 / $2.00 ngati mtundu wina uliwonse wa mlingo woyambira umene timatumikira. Kusinthanaku ndi kopapatiza ndipo kosavuta kulingalira: mumasiya zotuluka zokonzedwa zokakamizidwa ndi seva ndipo mumasunga zina zonse.

01Kodi Laguna S 2.1 ndi chiyani?

Laguna S 2.1 ndi mtundu wa zolemera zotseguka wotulutsidwa ndi Poolside pa July 21, 2026 pansi pa layisensi yololera ya OpenMDW-1.1 — membala wapakati wa banja la Laguna, pakati pa Laguna XS 2.1 (33B/3B zogwira ntchito) ndi Laguna M.1 (225B/23B zogwira ntchito).

Kapangidwe kofalitsidwa ndi router ya token-choice yokhala ndi softplus gating pa akatswiri 256 ogawidwa ntchito kuphatikiza katswiri mmodzi wogawana, khumi osankhidwa pa token iliyonse, pa zosanjikiza 48 — khumi ndi ziwiri za global-attention, makumi atatu ndi zisanu ndi chimodzi za sliding-window pa zenera la ma token 512, kusinthasintha kumene kumasunga mtengo wa chisamalira wa zolowetsa zazitali kwambiri pafupi ndi linear m'malo mwa quadratic. Chisamalira ndi grouped-query chokhala ndi mitu 8 ya key-value pa mtundu wa 128.

Zotsatira ziwiri ndizofunika kwa wogwiritsa ntchito API. Kuyatsa kosakanikirana ndicho chifukwa mtundu wa 118B ungathe kukhala pa mtengo wotsika kwambiri — zigawo ~8B zokha ndizo zimagwira ntchito pa token, choncho masamu pa token ali pafupi ndi mtundu waung'ono wolimba kuposa mmene chiwerengero cha zigawo chikuwonetsera. Ndipo kusinthasintha kwa sliding-window ndicho chifukwa zenera lalitali ndi logwiritsika ntchito zoonadi m'malo mokhala lokha pa mawu. Kutulutsidwa kwa Poolside kumathandiza ma token 1,048,576; timatumikira 262,144, kufanana ndi mtundu wina uliwonse pa gateway.

02Chodziwitsa chimodzi: palibe zotuluka zokonzedwa

Chomwe chikadali chikugwira ntchito. Kuyitanitsa ntchito kumathandizidwa kwathunthu, ndipo ma parameter a kuyitanitsa chida amafikabe ngati chinthu cha JSON ndendende ngati pa mtundu wina uliwonse pano. Ngati "zotuluka zokonzedwa" za chinthu chanu ndi ma parameter a chida zenizeni — zoona kwa machitidwe ambiri a agent, maseva a MCP ndi ma database front-end — Laguna S 2.1 imakuthandizani bwino. Streaming imagwira ntchito pa zilankhulo zonse zitatu; reasoning_effort imagwira ntchito.

Chomwe sichikugwira ntchito. Simungathe kupatsa endpoint JSON Schema ndikutsimikizidwa kuti yankho la mawu la wothandizira likutsimikizika motsutsana nayo, ndipo response_format: {"type": "json_object"} sikukupatsani kalikonse. Kufunsa JSON nthawi zambiri kumatulutsa JSON — uwu ndi mtundu wodziwa ntchito — koma "nthawi zambiri" si mgwirizano. Poolside yokha imalemba za kulakwitsa kwa JSON escaping mu mayitanidwe a chida ozama ngati malire odziwika, chikumbutso chakuti chitsimikizo, kumene chilipo kwina mu mndandanda wathu, chimachokera ku chosanjikiza chotumikira m'malo mwa zolinga zabwino za mtunduwo.

Mnzake uti wogwiritsa ntchito m'malo mwake

Ngati ntchito yanu ikufunika JSON yotsimikizika, musamenyane ndi mtunduwu — sinthani id ya mtundu. Anzake anayi ali pa mtengo womwewo wa $0.50 / $2.00 ndi chithandizo chonse cha json_schema ndi response_format:

Id ya mtundu	json_schema	Masomphenya	Chifukwa chomwe mungasankhe
MiMo-V2.5-Pro-W8A16	Inde	Ayi	Kusintha kolunjika — zolemera za 8-bit, mlingo wolondola kwambiri, mtengo womwewo
MiMo-V2.5-W8A16	Inde	Inde	Chomwecho, kuwonjezera zolowetsa za zithunzi
Hy3-W8A16	Inde	Ayi	Mtundu wina wa 8-bit wa mawu pa mtengo wotsika kwambiri
DeepSeek-V4-Flash-0731	Inde	Ayi	W4A16-AutoRound-REAP ngati Laguna, chithandizo cha schema chilipo

Chosankha chachizolowezi chabwino ndi MiMo-V2.5-Pro (W8A16): mawu okha, mtengo wotsika kwambiri, wopanikizidwa pa mlingo wofewa kwambiri umene timayendetsa — tsatanetsatane pa kaadi ya mtundu ya MiMo-V2.5. Chifukwa mitundu yonse khumi ndi iwiri imagawana ma endpoint, context ndi malire a kutuluka, kusinthira ndi kusintha kwa mzere umodzi ku munda wa model.

03Kodi W4A16-AUTOROUND-REAP amatanthauza chiyani?

Chowonjezera pa id ya mtundu si chokongoletsa — chimanena ndendende mmene zolemera zinapanikizidwira zisanaikidwe pa gulu lathu la ma GPU. Zigawo zitatu.

W4A16 — zolemera za 4-bit, zochitika za 16-bit

Zolemera zimasungidwa pa mabiti 4; zochitika — mtengo umene umayenda pakati pa zosanjikiza pa nthawi ya kuganiza — zimakhalabe pa 16. Kudula zolemera kufika gawo limodzi mwa anayi la malo awo a BF16 ndicho chimapangitsa mtundu wa 118B kukwana ndi kupitiriza kufulumira, pomwe kusunga zochitika zotakata kumateteza gawo lodalira manambala la kudutsa kwakutsogolo. Panikizani chomwe ndi chachikulu ndi chosasintha, siyani chomwe ndi chaching'ono ndi chosinthasintha.

AutoRound — kuzungulira kophunziridwa, osati kuzungulira ku pafupi

Njira yosaganizira ya quantization ndi kuzungulira ku pafupi kwambiri: sinthani kulemera kulikonse ku mtengo woyandikira kwambiri wa 4-bit, vomerani cholakwika. AutoRound, chida chotseguka cha Intel cha quantization pambuyo pa maphunziro, chimachita bwino kuposa: njira yake ya SignRound imagwiritsa ntchito signed gradient descent kukonza pamodzi zisankho za kuzungulira ndi malire a kudula pa masitepe angapo mazana, kusankha pa kulemera kulikonse ngati kuzungulira pamwamba kapena pansi kumawononga pang'ono zotulutsa zenizeni za chosanjikiza. Chimatenga lingaliro kuchokera ku quantization-aware training popanda mtengo wa kuphunzitsanso, ndipo ndicho chifukwa build ya W4A16 ingakhale mtundu weniweni m'malo mokhala chodabwitsa chotayika.

REAP — kudula akatswiri m'malo mowaphatikiza

REAP ndi Router-weighted Expert Activation Pruning, yofalitsidwa ndi Cerebras Research. Mu mtundu wa MoE si akatswiri onse amagwira ntchito yofanana: router imatumiza ma token ochepa kwa ena, ndipo ena amapereka gawo laling'ono pamene akugwira ntchito. REAP imayerekezera katswiri aliyense pa zizindikiro ziwiri — mfundo za router gate ndi ma norm a kugwira ntchito kwa katswiri — ndipo imachotsa ofunika pang'ono mu kudutsa kamodzi, quantization isanachitike. Chofunika kwambiri ndi kuti imadula m'malo mophatikiza: kuphatikiza kumasakaniza akatswiri angapo kukhala mmodzi, kuwononga ulamuliro wolondola, wodalira zolowetsa wa router ndikuyambitsa pansi pa cholakwika chimene kukonza kulikonse sikungachotse (olemba amatcha izi functional subspace collapse), pomwe kudula kumasiya opulumuka ndi ulamuliro wa router pa iwo wosasinthidwa. Cerebras ikunena kuti kuchotsa mpaka 50% ya akatswiri kuchokera ku mitundu ya tiriliyoni ya zigawo pomwe kukusunga mtundu wa maziko kwambiri — pa mtundu wa coder wa 480B pa kudula kwa 50%, 97.6% ya kuthekera kwa maziko kwa code yosakhala agentic ndi 96.7% pa agentic SWE-Bench. Izo ndi ziwerengero za Cerebras za mitundu yawo yokha, osati miyeso ya build yathu, koma ndizo chifukwa timayendetsa chomangamanga ichi pa maziko amene mphamvu yake yaikulu ndi code.

Pamodzi chowonjezerachi chimawerengeka ngati chiganizo: dulani akatswiri amene sanagwiritsidwe ntchito kwenikweni, phunzirani kuzungulira kwa 4-bit m'malo moyerekeza, sungani zochitika pa kutakata kwathunthu.

04Kodi mtunduwu ndi wabwino kwenikweni pa chiyani?

Choona poyamba: Laguna S 2.1 ndi mtundu wa kambirano wa mawu okha — palibe zolowetsa za masomphenya pa id iyi ya mtundu — ndipo Poolside inaupanga kuti ukhale wa code ya agentic ndi ntchito zautali, osati mawu. Maphunziro apambuyo pake anayenda pa malo 409,000, kuphatikiza ntchito 83,000 za terminal ndi 168,000 za software-engineering, ndi reinforcement learning mu FP8. Zotsatira zofalitsidwa ndi Poolside za kutulutsidwa kosapanikizidwa:

Benchmark	Kuganiza	Popanda kuganiza	Chomwe chimayeza
SWE-Bench Multilingual	78.5%	—	Kukonza zolakwika zenizeni pa zilankhulo zambiri
Terminal-Bench 2.1	70.2%	60.4%	Ntchito ya shelo ndi zida za masitepe ambiri
SWE-Bench Pro	59.4%	—	Ntchito zovuta kwambiri, zopewa contamination
Toolathlon Verified	49.7%	—	Kutakata kwa kugwiritsa ntchito zida pa ma API ambiri
SWE Atlas	46.2%	—	Kuyenda mu chosungiramo chautali
DeepSWE v1.1	40.4%	16.5%	Uinjiniya wa mapulogalamu wa agentic

Mpata pakati pa zigawo ziwirizi ndicho nambala yothandiza kwambiri pano: pa DeepSWE chiwerengero chimakwera kuwirikiza kuposa kaŵiri pamene kuganiza kwayatsidwa, ndipo pa Terminal-Bench chimapindula pafupifupi mfundo khumi.

Ziwerengero zofalitsidwa ndi Poolside za kutulutsidwa kwa Laguna S 2.1 kosapanikizidwa, July 2026. Si miyeso ya build yathu ya W4A16-AutoRound-REAP, ndipo sitikufalitsa manambala athu okha a benchmark pa iyi. Gwiritsani ntchito reasoning_effort kukonza ndalama za kuganiza pa API yathu.

Pa ntchito ya kambirano ndi kupanga malingaliro, mfundo yoona ndi iyi: makhalidwe amene ali kumbuyo kwa kuyenda kwautali kwa agentic — kusunga mkhalidwe wambiri mogwirizana, kukhalabe pa ntchito, kusapatuka — ndi omwewo amene ali kumbuyo kwa ntchito yolemba munthu wautali, ndipo mu ntchito zenizeni imasunga mawu bwino pa zenera lalitali kwambiri. Palibe benchmark yofalitsidwa yomwe ikutero, ndipo sitidzapanga imodzi. Yesani pa mafunso anu; pa $0.50 pa miliyoni ya ma token olowa izi ndi pafupifupi zaulere.

05Kodi mtengo wotsika kwambiri umagula chiyani kwenikweni?

$0.50 kulowa / $2.00 kutuluka pa miliyoni ya ma token ndi mlingo woyambira pa Shannon AI Gateway, wogawidwa ndi MiniMax-M3, DeepSeek-V4-Flash, MiMo-V2.5, MiMo-V2.5-Pro ndi Hy3. Zenizeni:

Ntchito	Kulowa	Kutuluka	Mtengo
Kutembenuka kumodzi kwa chat ndi mbiri ya 100K	100,000	2,000	$0.054
Mutu wa mawu 5,000 kuchokera ku lipoti lalitali	30,000	7,000	$0.029
Zotembenuka 1,000 zaifupi za kambirano	4,000,000	500,000	$3.00
Kuwerenga context yonse ya 262K	262,144	16,384	$0.164

Mzere womaliza ndi umene uyenera kukumbukiridwa: kudzaza zenera lonse la context ndikupanga yankho lalikulu kwambiri la ma token 16,384 kumalipiritsa pafupifupi masenti khumi ndi asanu ndi limodzi. Kudyetsa bukhu lonse, zolembedwa za kutembenuka, kapena gulu la zikalata mu yitanidwe limodzi kumasiya kukhala chisankho cha ndalama ndikukhala chisankho cha kuchedwa. Masamu omwewo amayendetsa ntchito ina ya mtengo wotsika kwambiri — kuchuluka: kuika m'magulu ya bulki, kulembanso kwa bulki, kambirano yopeka, kupanga koyamba komwe mtundu wamphamvu kwambiri umakonzanso pambuyo pake. Onani tsamba la mitengo ndi mndandanda wa kafukufuku.

06Chomwe "chosasidwa" chimatanthauza pano

Chimatanthauza chimene chimanena: palibe wosanjikiza wa kukana ndipo palibe kusefa kwa zinthu pa zotulutsa. Palibe chowunika chimene chili pakati pa mtunduwo ndi yankho lanu kusankha ngati mwaloledwa kufunsa; mtunduwo umayankha mu mtundu umene munafunsa ndipo sichikulangizani zoyambira poyamba.

Pa ntchito yopanga malingaliro icho ndicho kusiyana pakati pa chida chogwiritsika ndi chosagwiritsika. Nkhani zopeka zimakhala ndi anthu oopsa, anthu amene amanama mokopa, chiwawa chokhala ndi zotsatira, olongosola opanda makhalidwe abwino — mtundu umene umaduka mkhalidwe kuti uwonjezere chidziwitso chotchinjiriza walephera ntchitoyo, sanapambane pa chitetezo. Chomwecho chimagwira pa kafukufuku wa chitetezo, zolemba za red-team, kulemba kwa zachipatala ndi zamalamulo, ndi kutembenuza kumene gwero limakhala lachindunji: magulu kumene kukana ndi yankho lolakwika lokha. Chomwe sichitanthauza ndi kuti chiweruzo chimasiya kukhala chanu — mumakhalabe ndi udindo pa chomwe mukupanga, pansi pa Ndondomeko yathu ya Kagwiritsidwe Kaudindo ndi malamulo ogwira ntchito.

07Momwe mungayitanire Laguna API yosasidwa

Zilankhulo zitatu, mtundu umodzi: /v1/chat/completions (OpenAI), /v1/messages (Anthropic), /v1/responses. Streaming imagwira ntchito pa zonse zitatu, choncho kaphatikizidwe kamene kalipo kamasunthira posintha base URL ndi id ya mtundu — palibe kusintha kwa SDK, palibe kulembanso kwa kasamalidwe ka mauthenga anu.

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"
}

Zolemba zochepa zothandiza:

• Musatumize response_format kapena json_schema ku id iyi ya mtundu — ngati laibulale yanu ya kasitomala imaika imodzi mwazi mwachizolowezi, chotsani kapena sinthirani ku mnzake amene amaithandiza.
• Zida zimapita mu mndandanda wamba wa tools — kuyitanitsa ntchito kumathandizidwa kwathunthu; kukakamizidwa kwa schema ya mawu aulere kokha ndikosathandizidwa.
• Konzani ndalama za kuganiza kwanu. reasoning_effort imalemekezedwa — yotsika kapena yozimitsidwa imasunga kuchedwa kwa kambirano kotsika ndi ma token otuluka otsika mtengo; ikwezeni pa ntchito za zida za masitepe ambiri, kumene mpata wa benchmark pamwambapa ukuwonetsa chimene umagula.
• Malire a kutuluka ndi ma token 16,384 pa pempho. Pa kupanga kokulirapo bukhu, gawani m'mitu ndikutumiza mu streaming.
• Mawu okha. Palibe zolowetsa za zithunzi pa id iyi ya mtundu — gwiritsani ntchito MiMo-V2.5, Kimi-K2.6, Kimi-K3, MiniMax-M3 kapena inkling pa masomphenya.

Mndandanda wonse wa ma parameter, kutsimikizira, tanthauzo la zolakwika ndi malo oyeserera amoyo: zolemba za API.

08Nthawi yosankha ndipo nthawi yosasankha

Sankhani pamene zotulutsa zanu ndi mawu, kambirano kapena code yowerengedwa ndi munthu; mukufunika kuyitanitsa ntchito koma osati mawu okakamizidwa ndi schema; mtengo pa token ukulamulira kuposa context yaitali; kapena mukufuna maganizo ena kuchokera ku mzere wina kupatula mabanja a DeepSeek, Kimi ndi MiMo amene amadzaza mndandanda wina.

Sankhani chinanso pamene mukufunika JSON yotsimikizika (MiMo-V2.5-Pro), kumvetsetsa zithunzi (Kimi-K2.6 kapena MiMo-V2.5), kapena malire a kuvuta kumene mpata wa mtengo si wofunika poyerekeza ndi kupeza yankho lolondola (DeepSeek-V4-Pro, Kimi-K3).

09Mafunso ofunsidwa kawirikawiri

---

Magwero: Poolside — kaadi ya mtundu ya Laguna S 2.1 · Kudziwitsa Laguna S 2.1 · REAP the Experts (Cerebras Research) · Intel AutoRound. Ziwerengero za kapangidwe ndi benchmark zafalitsidwa ndi wofalitsa za kutulutsidwa kosapanikizidwa, osati miyeso ya build imene timatumikira. Shannon AI imayendetsedwa ndi Shannon Lab LLC, New Mexico, USA.