API na GLM marar tantancewa: mafi kyawun daraja a kowane token

Yawancin kwatancen "samfuri mafi rahusa" ba na gaskiya ba ne, saboda zaɓin mai rahusa yawanci ƙaramin samfuri ne sanye da tallar babban samfuri. Ba haka abin yake faruwa a nan ba. GLM-5.2 yana ɗaya daga cikin manyan tushe masu buɗaɗɗen nauyi da kowa ya taɓa fitarwa, kuma a $0.73 shiga / $2.34 fita yana zaune mataki ɗaya sama da ƙasan katalogin mu alhali yana ɗauke da tushe wanda ya ninka girman abin da wannan ƙasan yake saya sau da yawa. Wannan giɓi — babban samfuri, farashi kusa da ƙasa — shi ne dalilin da ya sa wannan shafin yake nan.

01Daga ina GLM ya fito: Tsinghua, cika wuraren da aka bari, da juyi mai ƙarfi zuwa wakilai

GLM ba wani abu ne da aka ɗauko daga Llama aka sake masa suna ba. Yana ɗaya daga cikin ƴan dangin samfura da suke da manufar horo ta kansu, wadda ƙungiyar bincike guda ta yi ta gyara ta shekaru biyar. Sunan yana nufin General Language Model, kuma takardar kafuwa — GLM: General Language Model Pretraining with Autoregressive Blank Infilling, ACL 2022 — ta gabatar da zaɓi na uku kusa da tsarin harshe mai rufaffiyar kalma na BERT da tsarin autoregression mai tsauri daga hagu zuwa dama na GPT. GLM yana share sassa masu haɗe sannan ya sake samar da su ta autoregression a jerin da aka gauraya, da rufa-rufan matsayi masu girma biyu waɗanda suke bin diddigin matsayin token a cikin takardar da aka ɓata da kuma matsayinsa a cikin sashen da ake cikawa: gine-gine guda yana ɗaukar fahimta da samarwa daidai da kyau.

Ƙungiyar da ke bayansa ita ce Knowledge Engineering Group ta Jami'ar Tsinghua, wadda ta haifi kamfanin da yanzu aka sani da Z.ai (a dā Zhipu AI). Aikinta na farko da ya yi suna shi ne GLM-130B a Oktoba 2022 — samfuri mai kauri na harsuna biyu Ingilishi/Sinanci mai sigogi biliyan 130, an fitar da nauyi da lambar a bainar jama'a a lokacin wallafawa. Fitarwa a buɗe ita ce salon gidan tun daga lokacin.

Zamanin MoE, da juyi zuwa wakilai

Zaren da ya ratsa duka ya fi kowane saki guda muhimmanci. Tun GLM-4.5, an inganta kowace sigar don abu ɗaya: ci gaba da haɗin kai a kan kiran kayan aiki da dama. Wannan manufa ce dabam da ƙara makin ma'auni na zagaye guda, kuma tana bayyana a matsayin ƙarancin gardamomin kayan aiki da suka lalace, ƙarancin karkacewa a cikin dogon shiri, da ƙarancin ɗabi'ar "ya sake fara aikin da ƙarfin gwiwa a zagaye na 14" wadda take sa tsarin wakilai ya yi tsada.

02Menene GLM-5.2, kuma a kan me aka auna shi

Z.ai yana wallafa lambobin ma'auni na GLM-5.2 a ma'ajiyar GLM-5. Biyu daga cikinsu suna daidaita da ainihin aikin wakilai maimakon tunawa da ƙananan bayanai:

Terminal-Bench 2.1 — ayyukan wakili na layin umarni masu dogon zango. Z.ai ne ya bayar da rahoto.

SWE-bench Pro — warware matsala a ainihin ma'ajiyar lamba. Z.ai ne ya bayar da rahoto; GLM-5.1 ya samu 58.4 a jarabawa guda.

Waɗannan lambobin dakin binciken ne, ba namu ba, kuma suna kwatanta tushen marar matsawa. Suna nuna maka nau'in samfurin da kake mu'amala da shi, ba garanti ba.

Bayani guda ya fi ma'auni daraja, saboda ya fito daga mai kimantawa da ba shi da wani abin da zai sa ya yabi samfurin. A watan Yuli 2026 Cibiyar Ma'auni da Sabbin Abubuwa ta AI ta Cibiyar Ƙasa ta Amurka don Ma'auni da Fasaha ta kimanta GLM-5.2 sannan ta kammala cewa wataƙila shi ne samfurin AI mai buɗaɗɗen nauyi mafi iyawa da ake da shi lokacin da aka fitar da shi, da iyawa gaba ɗaya kwatankwacin GPT-5.2 da iyawar tsaron yanar gizo kwatankwacin Claude Opus 4.6. Wannan kimantawar ta kuma lura da cewa GLM-5.2 yana toshe buƙatu masu laushi kaɗan idan aka kwatanta da samfuran tunani na Amurka, kuma ana iya kewaye shingayen da aka gasa cikin buɗaɗɗen nauyi da zarar an ɗauki wannan nauyin aka gudanar da shi da kai. Sashe na 05 zai koma kan hakan.

03Menene ma'anar "3BIT-REAP", da abin da yake kashe maka

REAP: datse masana maimakon haɗa su

Samfurin MoE mai sigogi 744B yana da masana 256 a kowace shimfida, waɗanda 8 daga cikinsu ke kunnuwa a kowane token. Router ɗin ba ya rarraba aiki daidai wa daida — ana kai wa wasu masana koyaushe, alhali ana kai wa dogon jela da wuya kuma ba sa bayar da gudummawa sosai idan an kai gare su. REAP (Router-weighted Expert Activation Pruning), daga Cerebras Research kuma an karɓe shi zuwa ICLR 2026, yana amfani da hakan: yana ba kowane masani maki bisa ma'aunin muhimmanci wanda ya haɗa ƙarfi da yawan lokutan da router yake ƙyale shi da girman gudummawar fitarwarsa, sannan ya cire waɗanda suka samu ƙarancin maki har abada.

Abin sha'awa shi ne kwatancin da REAP ya ci nasara a kai. Zaɓin bayyananne maimakon cire masani shi ne a haɗa shi cikin wanda ya tsira, wanda yake kama da mafi aminci. Abin da takardar ta gano a kan samfuran MoE daga sigogi 20B zuwa 1T shi ne cewa haɗawa a zahiri ta fi muni ga aikin samarwa, kuma datsawa tana tsayawa da kyau ƙwarai: kusan babu asara a kan ma'aunan samar da lamba da kiran kayan aiki ko bayan an cire rabin masanan, a kan tushe da suka haɗa da Qwen3-Coder-480B da Kimi-K2. Rayuwar lamba da kiran kayan aiki lafiya daidai shi ne abin da kake son a kiyaye a kan samfuri wanda dalilin wanzuwarsa shi ne injiniyancin wakilai.

Nauyi na 3-bit, da cinikin gaskiya

Bayan datsawa, ana ajiye nauyin da ya tsira a bit 3 — mai tsauri fiye da matakan W4A16-AUTOROUND-REAP da W8A16 a wasu wuraren katalogin mu, kuma an ajiye shi ga manyan tushe kaɗai, inda babu wata hanya ta sa samfuri mai sigogi kwata uku na tiriliyan a kan na'urar hidima a ci gaba da fitar da tokens a saurin da za a iya amfani da shi.

Matsawa ba kyauta ba ce. Hoton tunani daidai shi ne: datsawa da matsawa suna lalata dogon jela da farko. Ɗabi'un da ake yawan amfani da su kuma suke da kauri — bin umarni, tsarin lamba, tsara kiran kayan aiki, shirin matakai da yawa — suna tsira da kyau, wanda daidai shi ne abin da sakamakon REAP ya nuna. Tunawa da bayanan da ba a saba da su ba, harsuna da ba kasafai ba, da ƙananan bayanan fannoni na musamman su ne inda ya kamata ka sa ran bambanci idan aka kwatanta da GLM-5.2 marar matsawa. Idan aikinka shi ne "san wani bayani da ba a saba da shi ba sarai," ka biya matakin mafi aminci. Idan kuwa shi ne "gudanar da shiri a kan kiran kayan aiki arba'in ba tare da ɓacewa daga hanya ba," wannan shi ne matakin da aka tsara hanyar dominsa.

04Batun daraja: nawa API na LLM marar tantancewa mai rahusa yake kashewa da gaske

Farashin jerin API da Z.ai da kansa ya wallafa ga GLM-5.2 shi ne $1.40 kowane tokens miliyan na shigarwa da $4.40 kowane miliyan na fitarwa. Farashinmu a kan tushe guda shi ne $0.73 / $2.34 — kimanin kashi 48% ƙasa a shigarwa da kashi 47% ƙasa a fitarwa. Wannan ragin shi ne abin da matsawar take saya, kuma yana zuwa da abu na biyu wanda ba za ka iya saya daga masu shi a kowane farashi ba: babu shimfidar ƙi.

A cikin katalogin mu, farashi haɗe shi ne lambar da take da muhimmanci. Ga shi a rabon shigarwa-zuwa-fitarwa na 3:1, kimanin yadda wakili mai kiran kayan aiki yake a aikace — mahalli da sakamakon kayan aiki masu yawa suna shiga, tokens kaɗan kwatankwacinsu suna fita.

Samfuri	Shiga / 1M	Fita / 1M	Haɗe 3:1	Gani
GLM-5.2 · 3BIT-REAP	$0.73	$2.34	$1.13	A'a
Nemotron3Ultra · 3BIT-REAP	$0.75	$3.30	$1.39	A'a
Kimi-K2.6 · W4A16-AUTOROUND-REAP	$0.78	$3.67	$1.50	Eh
DeepSeek-V4-Pro-0813 · 3BIT-REAP	$1.95	$3.90	$2.44	A'a
inkling · W4A16-AUTOROUND-REAP	$1.42	$6.07	$2.58	Eh
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	$7.65	Eh

Samfura shida suna zaune a ƙasan $0.50 / $2.00 — MiniMax-M3, DeepSeek-V4-Flash, Laguna-S-2.1, MiMo-V2.5, MiMo-V2.5-Pro da Hy3 — wato ƙimar haɗe ta $0.875. GLM-5.2 yana kashi 29% sama da wannan ƙasan kuma yana saya maka tushe na ajin 744B da saka hannun jari mai dorewa a ɗabi'ar wakilai a bayansa. Duk abin da farashinsa yake sama da shi yana kashe aƙalla kashi 23% ƙari a haɗe, kuma Kimi-K3 yana kashe sau 6.8 ƙari. Wannan ita ce hujjar daraja a jere guda: wuri mafi rahusa a katalogin da za ka sayi tushe na wakilai na matakin sama.

An mayar da shi zuwa aiki

• Cika mahalli sau ɗaya — tokens 262,144 na shigarwa: $0.19. An karanta cikakkiyar babbar ma'ajiyar lamba cikin tagar a ƙasa da sisin ashirin.
• Kammalawa guda mai cikakken tsawo — tokens 16,384 na fitarwa: $0.038.
• Zagayen wakili guda — 20K shiga, 2K fita: $0.019. Zagaye guda yana kashe $0.115 a Kimi-K3 da $0.047 a DeepSeek-V4-Pro.
• Zagayen wakili ɗari: $1.93, idan aka kwatanta da $11.48 a samfuri mafi tsada a katalogin.

Ayyukan wakilai suna da nauyi a gefen shigarwa, kuma amfanin GLM ya fi girma daidai a can: farashin shigarwarsa yana ƙasa da na Kimi-K3 sau 5.2 kuma ƙasa da na DeepSeek-V4-Pro sau 2.7. Idan wakilinka yana sake aika tattaunawa mai girma a kowane zagaye — kuma kusan kowane wakili yana yin haka — wannan ninkin shi ne abin da yake ƙarewa a takardar biyanka.

05Menene "marar tantancewa" yake nufi a wannan ƙarshen hanyar

Abubuwa biyu, daidai: babu shimfidar ƙi — babu mai rarrabewa da yake yanke shawarar waɗanne buƙatu ne aka yarda samfurin ya gani, kuma babu samfurin manufa da yake sake rubuta umarninka kafin ya iso — da babu matatar fitarwa, wato tokens da samfurin ya samar su ne tokens da ka karɓa, ba a shafe su ba kuma ba a musanya su ba.

Wannan tsari ne dabam da API na kasuwanci da yake tallata hanyar "sassauƙa" ko ta "mai haɓakawa". Waɗancan har yanzu suna wuce da zirga-zirga ta tarin manufofin mai sayarwa; abin da ya canja shi ne ƙofar shiga, ba wanzuwar ƙofar ba. A nan babu ƙofar da za a daidaita — samfurin yana gudana a kan gungun GPU namu, kuma ɗabi'arsa ita ce ɗabi'ar tushen. Ka karanta binciken NIST CAISI daga sashe na 02 a matsayin bayanin fasaha maimakon gargaɗi sai ya bayyana maka kayan: samfuri wanda daidaitawarsa take zaune a cikin nauyinsa, ana hidimarsa ba tare da shimfidar manufa ta waje ba, yana nuna hali bisa nauyinsa.

Me ya sa wannan yake da muhimmanci ga wakilai musamman

Yawanci ana tattauna ƙi a matsayin batun abun ciki. A aikin wakilai galibi batun dogaro ne. Ƙi a zagaye na 31 na gudu mai zagaye 40 ba wata ƙin karɓa mai ladabi ba ce da za ka karanta ka wuce — mataki ne da ya lalace, mai tsara aiki wanda bai san yadda zai ci gaba ba, shiri da ya karye, da zagaye 30 na tokens da aka kashe ba tare da wani abin nunawa ba. Binciken tsaro, nazarin manhajar ɓarna, gwajin shiga, aikin ƙungiyar ja, almara ta manya, tsara takardun likita da na shari'a, da samar da bayanai don horon tsaro duk suna cin karo da wannan koyaushe, kuma gazawar tana da tsada daidai saboda tana faɗuwa a makare.

Ɗayan gefen shi ne shimfidar manufa yanzu taka ce. Amfani da Ƙofar Shiga ta Shannon AI yana ƙarƙashin Manufar Amfani da Alhaki tamu; cire hukuncin mai sayarwa daga zagayen yana nufin ka samar da naka.

06Abin da ƙarshen hanyar yake ba ka da gaske

Iyawa	GLM-5.2 · 3BIT-REAP
Alamar samfuri	GLM-5.2-3BIT-REAP
Taga mahalli	tokens 262,144
Matsakaicin fitarwa a kowace buƙata	tokens 16,384
Streaming	Eh
Kiran kayan aiki	Eh
Tunani / reasoning_effort	Eh
JSON Schema (json_schema)	Eh
response_format	Eh
Gani / shigar hoto	A'a — rubutu kaɗai
Shimfidar ƙi / matatar fitarwa	Babu

Ka karanta jeren gani a hankali kafin ka gina. GLM-5.2 rubutu kaɗai ne a ƙofar shigarmu. Idan kana buƙatar shigar hoto, katalogin yana da zaɓuɓɓuka biyar masu iya gani — MiniMax-M3 da MiMo-V2.5 a ƙasan $0.50 / $2.00, Kimi-K2.6 a $0.78 / $3.67, inkling a $1.42 / $6.07, da Kimi-K3 a $3.83 / $19.12. Salon aiki mai hankali shi ne a tura fahimtar hoto zuwa MiniMax-M3 sai kuma sauran komai zuwa GLM-5.2, wanda ba ya kashe komai fiye da canja zaren samfuri.

Yaruka uku, samfuri guda

Ana iya kaiwa ga samfuri guda ta hanyoyin buƙata guda uku, duka masu streaming: /v1/chat/completions (yaren OpenAI), /v1/messages (yaren Anthropic), da /v1/responses. Babu abin da ke canjawa a samfurin tsakaninsu — zaɓi wanda SDK ɗinka ya riga ya iya magana da shi, kuma yawanci canjin base URL guda ya rage tsakaninka da gudanarwa.

# OpenAI dialect — point your existing client at the gateway
from openai import OpenAI

client = OpenAI(
    base_url="https://api.shannon-ai.com/v1",
    api_key="sk-...",
)

resp = client.chat.completions.create(
    model="GLM-5.2-3BIT-REAP",
    messages=[{"role": "user", "content": "..."}],
    tools=tools,                 # tool calling supported
    reasoning_effort="medium", # low | medium | high
    stream=True,
)

Cikakken tsarin buƙata da amsa ga dukkan yaruka uku, tsarin kiran kayan aiki, siffofin abubuwan da suke faruwa a streaming, da filin gwaji mai mu'amala suna cikin takardun API na Shannon AI.

07Lokacin da za a zaɓi GLM-5.2 — da lokacin da ba za a zaɓe shi ba

Ka kai gare shi sa'ilin da

• Kana gudanar da wakilai. Dogayen zagayen kiran kayan aiki su ne abin da aka gina wannan zuriyar dominsa kuma aka yi mata horon baya akai-akai domin sa, kuma farashin shigarwa yana sa sake aika tattaunawa mai girma a kowane zagaye ya zama mai sauƙin biya.
• Kana aiki a kan babban tarin lamba. Mahalli na 262K, tsara kiran kayan aiki abin dogaro, da $0.19 don karanta dukkan tagar.
• Yawa ce ƙuntatawarka. Rarrabewa, cirar bayani, samar da bayanai na roba, sake rubutawa da yawa — duk inda kuɗin kowane kira ninki babban adadi yake yanke shawarar ko aikin zai yiwu.
• Ƙi yana lalata bututunka. Binciken tsaro, gwajin adawa, almara marar tantancewa, tsara takardu masu laushi — wannan gazawar ba ta nan.

Ka kai ga wani abu dabam sa'ilin da

• Kana buƙatar gani. Rubutu kaɗai. Duba zaɓuɓɓuka masu iya gani da ke sama.
• Kana buƙatar tunani mafi zurfi na zagaye guda. Ga matsalolin nazari mafi wuya na harbi guda, DeepSeek-V4-Pro-0813 · 3BIT-REAP shi ne kayan aiki mai nauyi a kan sau 2.2 na kuɗin haɗe — a kan kiraye-kirayen da rufin shi ne ƙuntatawa, wannan shi ne cinikin da ya dace.
• Kana buƙatar matsakaicin aminci a kan ilimi da ba a saba da shi ba. Matakin W8A16 — MiMo-V2.5-Pro da Hy3 — an matsa shi da ƙarancin tsauri sosai kuma yana zaune a ƙasan $0.50 / $2.00, a kan ƙananan tushe.

Z.ai ne ya bayar da rahoton lambobin ma'auni na GLM-5.2 (Terminal-Bench 2.1, SWE-bench Pro) don tushen marar matsawa; ana jiran maimaitawa mai zaman kanta. An nuna lambobin Claude Opus 4.8 da GPT-5.5 don kwatanci kawai. Farashin jerin API na GLM-5.2 na asali kamar yadda Z.ai ya wallafa, Satumba 2026.

08Tambayoyin da ake yawan yi

---

Majiyoyi: Ma'ajiyar GLM-5 ta Z.ai · Du et al., GLM: General Language Model Pretraining with Autoregressive Blank Infilling, ACL 2022 · Zeng et al., GLM-130B: An Open Bilingual Pre-trained Model · Lasby et al., REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression, ICLR 2026 · Kimantawar NIST CAISI ta GLM-5.2 na Z.ai, Yuli 2026. Dakin binciken ne ya bayar da rahoton makin ma'aunin tushen kuma na wucin gadi ne; suna kwatanta GLM-5.2 marar matsawa, ba wannan ginin da aka matsa ba. Lambobin farashi, mahalli da iyawa ƙimomin katalogin Shannon AI ne tun 5 ga Satumba, 2026.