API na Laguna Marar Tantancewa

Yawancin shafukan "samfuri mai rahusa" suna neman ka karɓi ƙarami, gajiyayyen tushe don musanya da ƙaramar lamba. Wannan ba haka ba ne. Laguna S 2.1 sakin Mixture-of-Experts mai raguwa ne na tsara ta yanzu, an horar da shi a kan H200 guda 4,096 kuma an fitar da shi cikin ƙasa da makonni tara, kuma yana zaune a ƙasan $0.50 / $2.00 guda kamar kowane samfurin matakin shiga da muke hidimarsa. Cinikin yana da kunkuntar kuma yana da sauƙin fahimta: ka bar fitarwa mai tsari da uwar garke ke tilastawa ka riƙe duk sauran.

01Menene Laguna S 2.1?

Laguna S 2.1 samfuri ne mai buɗaɗɗen nauyi da Poolside ya saki a 21 ga Yuli, 2026 ƙarƙashin lasisin OpenMDW-1.1 mai saukin kai — memba mai matsakaicin girma na dangin Laguna, tsakanin Laguna XS 2.1 (33B/3B masu kunnuwa) da Laguna M.1 (225B/23B masu kunnuwa).

Gine-ginen da aka wallafa router ne na zaɓin token mai ƙofar softplus a kan masana 256 da ake turawa gare su tare da masani guda da ake rabawa, ana zaɓar goma a kowane token, a kan shimfidu 48 — goma sha biyu na kulawa ta gaba ɗaya, talatin da shida na taga mai zamewa a taga na tokens 512, haɗin da yake riƙe kuɗin kulawa na dogayen shigarwa kusa da madaidaiciya maimakon ninki biyu. Kulawa ta ƙungiyar tambaya ce da kai 8 na maɓalli-da-daraja a girman kai 128.

Sakamako biyu suna da muhimmanci ga mai amfani da API. Kunnuwa masu raguwa su ne dalilin da ya sa samfuri mai 118B zai iya zama a farashi mafi ƙasa — kusan sigogi biliyan 8 kaɗai ne suke aiki a kowane token, don haka lissafi a kowane token ya fi kusa da ƙaramin samfuri mai yawa fiye da yadda adadin sigogi yake nunawa. Kuma haɗin taga mai zamewa shi ne dalilin da ya sa dogon taga yake da amfani da gaske maimakon suna kawai. Sakin Poolside yana goyon bayan tokens 1,048,576; muna hidimar 262,144, daidai da kowane samfuri a ƙofar shiga.

02Gargaɗi guda: babu fitarwa mai tsari

Abin da har yanzu yake aiki. Ana goyon bayan kiran kayan aiki gaba ɗaya, kuma hujjojin kiran kayan aiki har yanzu suna zuwa a matsayin abin JSON daidai kamar a kowane samfuri a nan. Idan "fitarwa mai tsari" ta manhajarka da gaske ita ce hujjojin kayan aiki — gaskiya ne ga yawancin tsarin wakilai, uwayen garken MCP da gaban tsarin bayanai — Laguna S 2.1 yana hidimar ka da kyau. Streaming yana aiki a kan yaruka ukun; reasoning_effort yana aiki.

Abin da ba ya aiki. Ba za ka iya ba ƙarshen hanyar JSON Schema ka kuma tabbata cewa amsar rubutu ta mataimakin za ta yi daidai da ita ba, kuma response_format: {"type": "json_object"} ba ya saya maka komai. Neman JSON a umarni yakan samar da JSON — wannan samfuri ne mai iyawa — amma "yakan" ba yarjejeniya ba ce. Poolside da kansa ya rubuta kuskuren tserewar JSON a cikin kiran kayan aiki masu zurfin nannaɗewa a matsayin iyaka da aka sani, tunatarwa cewa tabbacin, inda yake a wani wuri a jerinmu, yana zuwa daga shimfidar hidima maimakon daga kyawawan niyyar samfurin.

Wane ɗan'uwa za a yi amfani da shi maimakon haka

Idan aikinka yana buƙatar JSON da aka tabbatar, kada ka yi faɗa da wannan samfurin — ka canja alamar samfurin. Ƴan'uwa huɗu suna zaune a farashin $0.50 / $2.00 guda tare da cikakken goyon bayan json_schema da response_format:

Alamar samfuri	json_schema	Gani	Dalilin zaɓarsa
MiMo-V2.5-Pro-W8A16	Eh	A'a	Musanyawa kai tsaye — nauyi na 8-bit, matakin mafi girman aminci, farashi guda
MiMo-V2.5-W8A16	Eh	Eh	Guda, tare da shigar hoto
Hy3-W8A16	Eh	A'a	Wani samfurin rubutu na 8-bit a farashi mafi ƙasa
DeepSeek-V4-Flash-0731	Eh	A'a	W4A16-AutoRound-REAP kamar Laguna, goyon bayan schema yana nan

Zaɓi na asali na halitta shi ne MiMo-V2.5-Pro (W8A16): rubutu kaɗai, a farashi mafi ƙasa, an matsa shi a matakin mafi laushi da muke gudanarwa — cikakkun bayanai a kan katin samfuri na MiMo-V2.5. Saboda duka samfura goma sha biyu suna raba ƙarshen hanya, taga mahalli da rufin fitarwa, canjawa canjin layi guda ne ga filin model.

03Menene ma'anar W4A16-AUTOROUND-REAP?

Ƙarin da ke kan alamar samfurin ba ado ba ne — yana faɗin daidai yadda aka matsa nauyin kafin a ɗora shi a kan gungun GPU namu. Sassa uku.

W4A16 — nauyi na 4-bit, kunnuwa na 16-bit

Ana adana nauyi a 4 bits; kunnuwa — darajojin da suke gudana tsakanin shimfidu a lokacin tunani — suna zama a 16. Yanke nauyi zuwa kashi ɗaya cikin huɗu na sawunsa na BF16 shi ne abin da ya sa samfuri mai 118B ya shiga kuma ya ci gaba da sauri, alhali riƙe kunnuwa masu faɗi yana kare ɓangaren wucewa gaba mai laushi a lissafi. Ka matsa abin da yake babba kuma marar motsi, ka bar abin da yake ƙarami kuma mai motsi.

AutoRound — zagayen da aka koya, ba zagaye zuwa mafi kusa ba

Hanya mai sauƙi ta matsi ita ce zagaye zuwa mafi kusa: ka manne kowane nauyi zuwa darajar 4-bit mafi kusa, ka karɓi kuskuren. AutoRound, kayan aikin matsi bayan horo buɗaɗɗe na Intel, ya fi haka — hanyarsa ta SignRound tana amfani da gangarawar gradient mai alama don inganta hukunce-hukuncen zagaye da iyakokin yanke tare a kan ƴan matakai ɗari biyu, tana yanke shawara a kowane nauyi ko zagaye sama ko ƙasa yake yin ƙarancin barna ga ainihin fitarwar shimfidar. Yana aro ra'ayin daga horo mai lura da matsi ba tare da kuɗin sake horo ba, kuma shi ne dalilin da ya sa ginin W4A16 zai iya zama samfuri na gaskiya maimakon abin sha'awa mai asara.

REAP — datsa masana maimakon haɗa su

REAP shi ne Router-weighted Expert Activation Pruning, wanda Cerebras Research ya wallafa. A cikin samfurin MoE ba duka masana ne suke ja nauyinsu ba: router yana aika ƴan tokens zuwa wasu, kuma wasu suna bayar da ƙaramin girma idan sun kunna. REAP yana ba kowane masani maki a kan alamun biyun — darajojin ƙofar router da ma'aunin kunnuwan masani — kuma yana cire waɗanda suke da ƙarancin muhimmanci a harbi guda, kafin matsi. Muhimmanci shi ne yana datsawa maimakon haɗawa: haɗawa tana naɗe masana da dama zuwa guda, tana lalata ikon router mai daki-daki, mai dogaro ga shigarwa kuma tana gabatar da ƙasan kuskure da babu daidaitawa da za ta cire (marubutan suna kiran sa faɗuwar sararin aiki), alhali datsawa tana barin waɗanda suka tsira da ikon router a kansu ba tare da lalacewa ba. Cerebras ya bayar da rahoton cire har zuwa kashi 50% na masana daga samfura masu sigogi tiriliyan alhali ya kiyaye ingancin asali gaba ɗaya — a kan samfurin rubuta lamba mai 480B a datsawar 50%, 97.6% na iyawar rubuta lamba mara ta wakilai na asali da 96.7% a SWE-Bench na wakilai. Waɗannan lambobin Cerebras ne ga samfuransu, ba aunan ginin mu ba, amma su ne dalilin da ya sa muke gudanar da wannan tarin a kan tushen da babban ƙarfinsa lamba ce.

Tare ƙarin yana karantuwa kamar jimla: ka datse masanan da ba a fi amfani da su ba, ka koyi zagayen 4-bit maimakon ka yi tsammani, ka riƙe kunnuwa a cikakken faɗi.

04Wane abu ne wannan samfurin yake da kyau a kai da gaske?

Gaskiya da farko: Laguna S 2.1 samfurin tattaunawa ne na rubutu kaɗai — babu shigar gani a wannan alamar samfurin — kuma Poolside ya gina shi don rubuta lamba na wakilai da aiki mai dogon zango, ba don rubutu ba. Horo na baya ya gudana a kan muhalli sama da 409,000, haɗe da ayyukan tasha 83,000 da na injiniyan software 168,000, tare da koyo mai ƙarfafawa a FP8. Sakamakon da Poolside ya bayar da rahotonsu ga sakin da ba a matsa ba:

Ma'auni	Tunani	Babu tunani	Abin da yake aunawa
SWE-Bench Multilingual	78.5%	—	Ainihin gyaran kuskure a harsuna da dama
Terminal-Bench 2.1	70.2%	60.4%	Aikin tasha da kayan aiki mai matakai da dama
SWE-Bench Pro	59.4%	—	Ayyukan ma'ajiya masu wuya, masu jure gurɓata
Toolathlon Verified	49.7%	—	Faɗin amfani da kayan aiki a kan API da dama
SWE Atlas	46.2%	—	Zirga-zirgar ma'ajiya mai dogon zango
DeepSWE v1.1	40.4%	16.5%	Injiniyan software na wakilai

Giɓin tsakanin ginshiƙai biyun ita ce lambar da za ta fi taimaka maka a nan: a DeepSWE makin ya fi ninka lokacin da aka kunna tunani, kuma a Terminal-Bench ya samu kusan maki goma.

Lambobin da Poolside ya bayar da rahotonsu ga sakin Laguna S 2.1 da ba a matsa ba, Yuli 2026. Ba aunan ginin mu na W4A16-AutoRound-REAP ba ne, kuma ba ma wallafa lambobin ma'auni namu a kansa. Ka yi amfani da reasoning_effort don sarrafa kasafin tunani a kan API ɗinmu.

Ga aikin tattaunawa da na kirkira matsayin gaskiya shi ne: halayen da ke bayan dogayen gudun wakilai — riƙe yanayi mai yawa a haɗe, zama a kan aikin, rashin kaucewa — su ne halayen da ke bayan aikin hali mai dogon tsari, kuma a aikace yana riƙe murya da kyau a kan taga mai tsawo sosai. Babu ma'aunin da aka wallafa da ya ce haka, kuma ba za mu ƙirƙiro ɗaya ba. Ka gwada shi a kan umarninka; a $0.50 kowane tokens miliyan na shigarwa hakan kusan kyauta ne.

05Menene farashi mafi ƙasa yake saya da gaske?

$0.50 shiga / $2.00 fita kowane tokens miliyan shi ne matakin shiga a Ƙofar Shiga ta Shannon AI, ana rabawa da MiniMax-M3, DeepSeek-V4-Flash, MiMo-V2.5, MiMo-V2.5-Pro da Hy3. A takamaime:

Aiki	Shigarwa	Fitarwa	Kuɗi
Juyin hira guda da tarihin 100K	100,000	2,000	$0.054
Babi na kalmomi 5,000 daga dogon bayani	30,000	7,000	$0.029
Juyin tattaunawa gajeru 1,000	4,000,000	500,000	$3.00
Karatun cikakken mahalli na tokens 262K	262,144	16,384	$0.164

Layi na ƙarshe shi ne wanda za a haddace: cika duka taga mahalli da samar da matsakaicin amsar tokens 16,384 yana kashe kusan cent sha shida. Ciyar da duka rubutun littafi, rubutun zama, ko tarin takardu cikin kira guda ya daina zama hukuncin kasafin kuɗi ya zama hukuncin jinkiri. Lissafi guda yana tuƙa ɗayan amfani da farashi mafi ƙasa — yawa: rarrabewa mai yawa, sake rubutawa mai yawa, tattaunawar roba, tsarin farko da samfuri mafi ƙarfi zai tace daga baya. Ka duba shafin farashi da fihirisar bincike.

06Menene ma'anar "marar tantancewa" a nan

Yana nufin abin da yake faɗa: babu shimfidar ƙi kuma babu matatar abun ciki a kan fitarwa. Babu mai rarrabewa da yake zaune tsakanin samfurin da amsarka yana yanke shawarar ko an ba ka izinin tambaya; samfurin yana amsawa a cikin salon da ka nema kuma ba ya yi maka wa'azi a kan tushen tambayar tukuna.

Ga aikin kirkira wannan shi ne bambanci tsakanin kayan aiki mai amfani da mara amfani. Almara tana ɗauke da abokan gaba masu barazana, halayen da suke ƙarya cikin gaskatawa, tashin hankali mai sakamako, masu ba da labari da ɗabi'arsu ta lalace — samfurin da ya karya hali don ya ƙara sanarwar gargaɗi ya gaza aikin, bai yi nasara a tsaro ba. Haka yake ga binciken tsaro, takardun red-team, rubutun likita da shari'a, da fassarar inda tushen yake da tsauri: rukunan da ƙi kawai amsa mara kyau ce. Abin da ba ya nufi shi ne hukunci ya daina zama naka — kai ne har yanzu ke da alhakin abin da ka samar, ƙarƙashin Manufarmu ta Amfani da Alhaki da dokar da ta shafe ka.

07Yadda ake kiran API na Laguna marar tantancewa

Yaruka uku, samfuri guda: /v1/chat/completions (OpenAI), /v1/messages (Anthropic), /v1/responses. Streaming yana aiki a kan ukun, don haka haɗin da kake da shi yana motsawa ta canja URL na asali da alamar samfuri — babu musanya SDK, babu sake rubuta sarrafa saƙonninka.

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"
}

Ƴan bayanai masu amfani:

• Kada ka aika response_format ko json_schema zuwa wannan alamar samfurin — idan ɗakin karatun abokin cinikinka yana saita ɗaya ta asali, ka cire shi ko ka koma ga ɗan'uwa da yake goyon bayansa.
• Kayan aiki suna shiga cikin tsarin tools na yau da kullum — ana goyon bayan kiran kayan aiki gaba ɗaya; tilasta schema a kan rubutu kyauta kaɗai ba a goyon bayansa ba.
• Ka tsara kasafin tunaninka. Ana girmama reasoning_effort — ƙasa ko a kashe yana riƙe jinkirin tattaunawa ƙasa da tokens na fitarwa masu rahusa; ka ɗaga shi ga aikin kayan aiki mai matakai da dama, inda giɓin ma'auni na sama yake nuna abin da yake saya.
• Rufin fitarwa tokens 16,384 ne a kowace buƙata. Ga samarwa mai tsawon littafi, ka raba shi zuwa babuka ka yi streaming.
• Rubutu kaɗai. Babu shigar hoto a wannan alamar samfurin — ka yi amfani da MiMo-V2.5, Kimi-K2.6, Kimi-K3, MiniMax-M3 ko inkling don gani.

Cikakken bayanin sigogi, tabbatarwa, ma'anonin kuskure da wurin gwaji kai tsaye: takardun API.

08Yaushe za a zaɓe shi, kuma yaushe ba za a zaɓe shi ba

Ka zaɓe shi lokacin da fitarwarka rubutu ne, tattaunawa ko lamba da mutum yake karantawa; kana buƙatar kiran kayan aiki amma ba rubutu da schema ta ƙuntata ba; kuɗi a kowane token ya mamaye a kan dogon mahalli; ko kana son ra'ayi na biyu daga zuriya ban da dangin DeepSeek, Kimi da MiMo da suka cika sauran jerin.

Ka zaɓi wani abu daban lokacin da kana buƙatar JSON da aka tabbatar (MiMo-V2.5-Pro), fahimtar hoto (Kimi-K2.6 ko MiMo-V2.5), ko ƙarshen wahala na gaba inda giɓin farashi ba shi da muhimmanci a gaban samun amsar daidai (DeepSeek-V4-Pro, Kimi-K3).

09Tambayoyin da ake yawan yi

---

Majiyoyi: Poolside — katin samfuri na Laguna S 2.1 · Introducing Laguna S 2.1 · REAP the Experts (Cerebras Research) · Intel AutoRound. Mai wallafa ne ya bayar da rahoton lambobin gine-gine da ma'auni ga sakin da ba a matsa ba, ba aunan ginin da muke hidimarsa ba. Shannon Lab LLC, New Mexico, Amurka, ce take gudanar da Shannon AI.