API ya GLM Yosasidwa: mtengo wabwino kwambiri pa token
Maziko a GLM a agent a Z.ai, otumikiridwa popanda wosanjikiza wa kukana pa gulu lathu la ma GPU — context ya 262K, kuyitanitsa zida, ndi mtengo wosakanizidwa wotsika theka la mtengo umene mzera womwewo umalipiritsa kwa kampani yeniyeni.
Chidule
GLM-5.2-3BIT-REAP ndicho mtundu pa Shannon AI Gateway umene timalimbikitsa mwachikale. Umayendetsa maziko a Mixture-of-Experts a sparse a ma parameter 744-biliyoni kuchokera ku mzera wa GLM — ophunzitsidwa mwachindunji pa ntchito za zida za agent kuyambira GLM-4.5 — ofinyidwa ndi zolemera za 3-bit ndi kudula akatswiri kwa REAP kuti akwane ndikukhalabe othamanga. Mumapeza context ya ma token 262,144, ma token 16,384 a zotulutsa, kuyitanitsa zida, kulingalira ndi reasoning_effort yosinthika, ndi zotulutsa za JSON Schema, pa $0.73 / $2.34 pa ma token miliyoni. Palibe wosanjikiza wa kukana ndipo palibe chosefera cha zotulutsa. Mawu okha — palibe masomphenya.
Kuyerekeza kwambiri kwa "mtundu wotsika mtengo kwambiri" ndi kosakhulupirika, chifukwa chosankha chotsika mtengo nthawi zambiri ndi mtundu waung'ono wovala malonda a mtundu waukulu. Ichi si chimene chikuchitika pano. GLM-5.2 ndi imodzi mwa maziko akulu kwambiri a zolemera zapoyera amene wina aliyense watulutsa, ndipo pa $0.73 kulowa / $2.34 kutuluka ili sitepe imodzi pamwamba pa pansi pa katalogi yathu pamene ikunyamula maziko ochulukirapo kambiri ya kukula kuposa chimene pansipo nthawi zambiri chimagula. Kusiyanaku — mtundu waukulu, mtengo pafupi ndi pansi — ndicho chifukwa cha tsamba ili.
01Komwe GLM idachokera: Tsinghua, blank infilling, ndi kutembenukira kwakukulu ku ma agent
GLM si Llama yosinthidwa dzina lokha. Ndi limodzi mwa mabanja ochepa a mitundu okhala ndi cholinga chake choyambirira cha pretraining, chokonzedwanso kwa zaka zisanu ndi gulu lomwelo la kafukufuku. Dzinalo limaimira General Language Model, ndipo pepala loyambitsa — GLM: General Language Model Pretraining with Autoregressive Blank Infilling, ACL 2022 — linapereka njira yachitatu pambali pa masked language modeling ya BERT ndi autoregression yolimba ya kumanzere-kupita-kumanja ya GPT. GLM imachotsa magawo olumikizana ndipo imawapanganso mwa autoregression mu dongosolo losakanikirana, ndi ma positional encoding a mbali ziwiri otsatira malo a token mu chikalata chosokonezedwa ndi malo ake mkati mwa gawo limene likudzazidwa: kapangidwe kamodzi kogwira ntchito bwino pa kumvetsa ndi kupanga zonse ziwiri.
Gulu limene linali kumbuyo kwake linali Knowledge Engineering Group ya Tsinghua University, imene inatulutsa kampani yotchedwa tsopano Z.ai (poyamba Zhipu AI). Chinthu chake choyamba chachikulu chinali GLM-130B mu Okotobala 2022 — mtundu wa dense wa zinenero ziwiri Chingerezi/Chichaina wa ma parameter 130-biliyoni, zolemera ndi kodi zotulutsidwa poyera pa nthawi ya kufalitsidwa. Kutulutsa poyera kwakhala kalembedwe ka bungweli kuyambira pamenepo.
Nthawi ya MoE, ndi kutembenukira ku ma agent
Mzere waukulu ndi wofunika kuposa kutulutsidwa kulikonse kokha. Kuyambira GLM-4.5, mtundu uliwonse wakonzedwa bwino pa chinthu chimodzomwe: kukhalabe wogwirizana pakati pa maitanidwe a zida makumi ambiri. Icho ndi cholinga chosiyana ndi kukweza manambala a mayeso a kutembenuka kumodzi, ndipo chimaonekera ngati zolakwika zochepa za maudindo a zida, kusochera kochepa mu ndondomeko yayitali, ndi kuchepa kwa khalidwe la "kuyambanso ntchito ndi chidaliro pa kutembenuka kwa 14" limene limapangitsa ma framework a agent kukhala okwera mtengo.
02Chimene GLM-5.2 chili, ndi chimene chinayezedwa
Z.ai imasindikiza manambala a mayeso a GLM-5.2 mu repository ya GLM-5. Awiri amagwirizana ndi ntchito zenizeni za agent osati kukumbukira zochepa:
Terminal-Bench 2.1 — ntchito za agent za command-line za nthawi yayitali. Zonenedwa ndi Z.ai.
SWE-bench Pro — kuthetsa mavuto a repository enieni. Zonenedwa ndi Z.ai; GLM-5.1 idapeza 58.4 pa mayeso omwewo.
Amenewa ndi manambala a labu, osati athu, ndipo amafotokoza maziko osafinyidwa. Amakhazikitsa gulu la mtundu umene mukugwira nawo ntchito, osati chitsimikizo.
Chidziwitso chimodzi ndi chofunika kuposa mayeso, chifukwa chimachokera kwa woyeza wopanda chikhumbo chotamandira mtunduwo. Mu Julayi 2026, US National Institute of Standards and Technology's Center for AI Standards and Innovation idayeza GLM-5.2 ndipo idamaliza kuti mwina inali mtundu wa AI wa zolemera zapoyera wamphamvu kwambiri wopezeka pamene inatulutsidwa, ndi mphamvu zonse zofanana ndi GPT-5.2 ndi mphamvu za cyber zofanana ndi Claude Opus 4.6. Kuyeza komweko kudapeza kuti GLM-5.2 imaletsa mapempho ochepa ovuta kuposa mitundu ya US reference, ndipo kuti zitetezo zophikidwa mu zolemera zapoyera zingapewedwe mukachotsa zolemerazo kuti zizichitidwa nokha. Gawo la 05 likubwereranso pa izi.
03Chimene "3BIT-REAP" chikutanthauza, ndi chimene chikukupatsani mtengo
REAP: kudula akatswiri m'malo mowaphatikiza
Mtundu wa MoE wa ma parameter 744B uli ndi akatswiri 256 pa chigawo chilichonse, amene 8 amagwira ntchito pa token iliyonse. Router siigawa ntchito mofanana — akatswiri ena amagwiritsidwa ntchito nthawi zonse, ndipo mchira wautali umagwiritsidwa ntchito mosowa ndipo umapereka zochepa pamene ukugwiritsidwa. REAP (Router-weighted Expert Activation Pruning), kuchokera ku Cerebras Research ndipo idavomerezedwa ku ICLR 2026, imagwiritsa ntchito izi: imayeza katswiri aliyense mwa mfundo yophatikiza mmene router imamugwiritsira ntchito mwamphamvu ndi kawirikawiri ndi kukula kwa zopereka za zotulutsa zake, kenako imachotsa kwamuyaya omwe apeza manambala otsika.
Gawo losangalatsa ndi kuyerekeza kumene REAP imapambana. Njira ina yodziwikiratu yosiyana ndi kuchotsa katswiri ndikumuphatikiza mu wina wotsalira, chomwe chimamveka chotetezeka. Zomwe pepalali linapeza pakati pa mitundu ya MoE kuchokera ku 20B mpaka 1T ma parameter ndikuti kuphatikiza n'koipa kwenikweni pa ntchito zopanga, ndipo kudula kumakhalabe bwino modabwitsa: pafupifupi popanda kutayika pa kupanga ma kodi ndi mayeso a kuyitanitsa zida ngakhale pambuyo pochotsa theka la akatswiri, pa maziko kuphatikizapo Qwen3-Coder-480B ndi Kimi-K2. Ma kodi ndi kuyitanitsa zida kukhalabe kosakhudzidwa ndicho chimene mukufuna kusungidwe pa mtundu umene chifukwa chake chilipo ndi uinjiniya wa agent.
Zolemera za 3-bit, ndi kusinthana koona
Pamwamba pa kudula, zolemera zotsalira zimasungidwa pa ma bit 3 — zolimba kuposa milingo ya W4A16-AUTOROUND-REAP ndi W8A16 kwina mu katalogi yathu, ndipo zosungidwa pa maziko akulu kwambiri okha, kumene kulibe njira ina yofikitsira mtundu wa ma parameter oposa magawo atatu mwa anayi a thililiyoni pa node yotumikira ndikusunga ma token akuyenda pa muyeso wogwiritsika.
Kufinya sikopanda mtengo. Mtundu woyenera wa maganizo: kudula ndi quantization zimawononga mchira wautali poyamba. Machitidwe olimba, ogwiritsidwa ntchito kawirikawiri — kutsatira malangizo, kapangidwe ka kodi, mawonekedwe a maitanidwe a zida, kukonzekera kwa masitepe ambiri — zimapulumuka bwino, chomwe ndicho chimene zotsatira za REAP zikusonyeza. Kukumbukira zenizeni zosowa, zinenero zosowa ndi zambiri za magawo apadera ndicho kumene muyenera kuyembekezera kusiyana ndi GLM-5.2 yosafinyidwa. Ngati ntchito yanu ndi "kudziwa mfundo yosazolowereka bwinobwino," lipirirani mlingo wa kulondola kwapamwamba. Ngati ndi "kuyendetsa ndondomeko pakati pa maitanidwe makumi anayi a zida popanda kuchoka pa njira," umenewu ndiwo mlingo umene njirayi idapangidwira.
04Mlandu wa mtengo: chimene API ya LLM yosasidwa yotsika mtengo imalipiritsadi kuyendetsa
Mtengo wa API wosindikizidwa wa Z.ai yokha wa GLM-5.2 ndi $1.40 pa ma token miliyoni yolowa ndi $4.40 pa miliyoni yotuluka. Mtengo wathu pa maziko omwewo ndi $0.73 / $2.34 — pafupifupi 48% wotsika pa kulowa ndi 47% wotsika pa kutuluka. Chotsitsa mtengocho ndicho chimene kufinya kumagula, ndipo chimabwera ndi chinthu chachiwiri chimene simungachigule kwa kampani yeniyeni pa mtengo uliwonse: palibe wosanjikiza wa kukana.
Mkati mwa katalogi yathu yokha, mtengo wosakanizidwa ndicho nambala yofunika. Nawu pa chiyerekezo cha 3:1 pakati pa kulowa ndi kutuluka, pafupifupi mmene agent wa kuyitanitsa zida amawonekera pa ntchito — context yambiri ndi zotsatira za zida zikulowa, ma token ochepa poyerekeza akutuluka.
| Mtundu | Kulowa / 1M | Kutuluka / 1M | Zosakanizidwa 3:1 | Masomphenya |
|---|---|---|---|---|
| GLM-5.2 · 3BIT-REAP | $0.73 | $2.34 | $1.13 | Ayi |
| Nemotron3Ultra · 3BIT-REAP | $0.75 | $3.30 | $1.39 | Ayi |
| Kimi-K2.6 · W4A16-AUTOROUND-REAP | $0.78 | $3.67 | $1.50 | Inde |
| DeepSeek-V4-Pro-0813 · 3BIT-REAP | $1.95 | $3.90 | $2.44 | Ayi |
| inkling · W4A16-AUTOROUND-REAP | $1.42 | $6.07 | $2.58 | Inde |
| Kimi-K3 · 3BIT-REAP | $3.83 | $19.12 | $7.65 | Inde |
Mitundu isanu ndi imodzi ili pa pansi pa $0.50 / $2.00 — MiniMax-M3, DeepSeek-V4-Flash, Laguna-S-2.1, MiMo-V2.5, MiMo-V2.5-Pro ndi Hy3 — mtengo wosakanizidwa wa $0.875. GLM-5.2 ili 29% pamwamba pa pansipo ndipo imakugulitsani maziko a gulu la 744B ndi ndalama zokhazikika za khalidwe la agent kumbuyo kwake. Chilichonse chomwe mtengo wake uli pamwamba pake chimalipiritsa osachepera 23% mokulirapo mosakanizidwa, ndipo Kimi-K3 imalipiritsa 6.8× mokulirapo. Ndilo mfundo ya mtengo mu mzere umodzi: malo otsika mtengo kwambiri mu katalogi kugula maziko a agent apamwamba kwambiri.
Zosandulika kukhala ntchito
- Kudzaza context konse — ma token 262,144 olowa: $0.19. Repository yonse yaikulu yowerengedwa mu zenera pa ndalama zosakwana masenti makumi awiri.
- Kutulutsa kumodzi kwa kutalika kwapamwamba — ma token 16,384 otuluka: $0.038.
- Kutembenuka kumodzi kwa agent — 20K kulowa, 2K kutuluka: $0.019. Kutembenuka komweko kumalipiritsa $0.115 pa Kimi-K3 ndi $0.047 pa DeepSeek-V4-Pro.
- Kutembenuka kwa agent zana limodzi: $1.93, poyerekeza ndi $11.48 pa mtundu wokwera mtengo kwambiri mu katalogi.
Ntchito za agent zimalemera kwambiri ku kulowa, ndipo phindu la GLM ndilo lalikulu kwambiri pamenepo: mtengo wake wa kulowa uli 5.2× wotsika kuposa wa Kimi-K3 ndi 2.7× wotsika kuposa wa DeepSeek-V4-Pro. Ngati agent wanu amatumizanso kukambirana kokulirakulira nthawi iliyonse — ndipo pafupifupi agent aliyense amachita zimenezi — chowonjezerachi ndicho chimapezeka pa bilu yanu.
05Chimene "chosasidwa" chikutanthauza pa endpoint iyi
Zinthu ziwiri, mwachindunji: palibe wosanjikiza wa kukana — palibe chosankha chinthu chosankha mapempho amene mtunduwo aloledwa kuwaona, ndipo palibe mtundu wa ndondomeko wolembanso prompt yanu isanafike — ndi palibe kusefa kwa zotulutsa, kutanthauza kuti ma token amene mtunduwo umapanga ndiwo ma token amene mumalandira, osachotsedwamo ndipo osasinthidwa.
Ichi ndi chinthu chosiyana ndi API ya malonda yolengeza njira ya "relaxed" kapena "developer". Zimenezo zikadalinso kutumiza magalimoto kudzera mu ndondomeko za kampani; chimene chimasintha ndi malire, osati kukhalapo kwa chipata. Pano palibe chipata choti musinthe — mtunduwo umayenda pa gulu lathu la ma GPU, ndipo khalidwe lake ndilo khalidwe la maziko ake. Werengani zomwe NIST CAISI idapeza mu gawo la 02 ngati mawu a ukadaulo m'malo mochenjeza ndipo zimafotokoza chinthucho: mtundu umene kugwirizana kwake kumakhala mu zolemera zake, wotumikiridwa popanda wosanjikiza wa ndondomeko wakunja, umachita malinga ndi zolemera zake.
Chifukwa chake izi ndi zofunika kwa ma agent makamaka
Kukana nthawi zambiri kumakambidwa ngati funso la zomwe zili mmenemo. Mu ntchito za agent ndi funso la kudalirika makamaka. Kukana pa kutembenuka kwa 31 kwa kuyenda kwa kutembenuka 40 si kukana modekha komwe mumawerenga ndikupitiriza — ndi sitepe yolakwika, orchestrator imene sidziwa mmene ingapitirire, ndondomeko yosweka, ndi ma token 30 a kutembenuka amene agwiritsidwa ntchito popanda kanthu koonetsa. Kafukufuku wa chitetezo, kusanthula kwa malware, kuyesa kwa kuloŵa mu dongosolo, red-teaming, nkhani zopeka za akuluakulu, kulemba kwa zachipatala ndi zalamulo, ndi kupanga ma dataset a kuphunzitsa chitetezo zonse zimakumana ndi izi nthawi zonse, ndipo kulephera kumeneku ndikokwera mtengo makamaka chifukwa kumafika mochedwa.
Mbali ina ndikuti wosanjikiza wa ndondomeko tsopano ndi wanu. Kugwiritsa ntchito Shannon AI Gateway kumayendetsedwa ndi Ndondomeko yathu ya Kagwiritsidwe Kaudindo; kuchotsa chiweruzo cha kampani mu njirayo kumatanthauza kupereka chanuchanu.
06Chimene endpoint imakupatsanidi
| Mphamvu | GLM-5.2 · 3BIT-REAP |
|---|---|
| Id ya mtundu | GLM-5.2-3BIT-REAP |
| Zenera la context | ma token 262,144 |
| Zotulutsa zapamwamba pa pempho | ma token 16,384 |
| Streaming | Inde |
| Kuyitanitsa zida | Inde |
Kulingalira / reasoning_effort | Inde |
JSON Schema (json_schema) | Inde |
response_format | Inde |
| Masomphenya / zolowetsa za zithunzi | Ayi — mawu okha |
| Wosanjikiza wa kukana / kusefa kwa zotulutsa | Palibe |
Werengani mzere wa masomphenya mosamala musanamange. GLM-5.2 ndi ya mawu okha pa gateway yathu. Ngati mukufuna zolowetsa za zithunzi, katalogi ili ndi njira zisanu za masomphenya — MiniMax-M3 ndi MiMo-V2.5 pa pansi pa $0.50 / $2.00, Kimi-K2.6 pa $0.78 / $3.67, inkling pa $1.42 / $6.07, ndi Kimi-K3 pa $3.83 / $19.12. Njira yanzeru ndi kutumiza kumvetsa zithunzi ku MiniMax-M3 ndi zina zonse ku GLM-5.2, chomwe sichilipiritsa kanthu kupatula mzere wa dzina la mtundu.
Zilankhulo zitatu, mtundu umodzi
Mtundu womwewo umafikika kudzera mu mawonekedwe atatu a pempho, onse streaming: /v1/chat/completions (chilankhulo cha OpenAI), /v1/messages (chilankhulo cha Anthropic), ndi /v1/responses. Palibe chokhudza mtunduwo chimene chimasintha pakati pawo — sankhani chomwe SDK yanu imalankhula kale, ndipo nthawi zambiri muli sitepe imodzi yosintha base-URL kuti muyende.
# 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, )
Ma schema onse a pempho ndi yankho pa zilankhulo zonse zitatu, mawonekedwe a kuyitanitsa zida, mawonekedwe a zochitika za streaming, ndi malo oyeserera amoyo ali mu zolemba za API za Shannon AI.
07Nthawi yoyenera kusankha GLM-5.2 — ndi nthawi yosayenera
Igwiritseni ntchito pamene
- Mukuyendetsa ma agent. Maloop aatali a kuyitanitsa zida ndicho chimene mzera uwu unamangidwira ndipo unaphunzitsidwiranso mobwerezabwereza, ndipo mtengo wa kulowa umapangitsa kukambirana kokulirakulira kukhala kotsika mtengo kuti mutumizenso nthawi iliyonse ya kutembenuka.
- Mukugwira ntchito mu codebase yaikulu. Context ya 262K, mawonekedwe odalirika a kuyitanitsa zida, ndi $0.19 kuwerenga zenera lonse.
- Kuchuluka ndicho vuto lanu. Kuika m'magulu, kutulutsa, kupanga data yopeka, kulembanso zambiri pamodzi — chilichonse chimene mtengo pa maitanidwe kuchulukitsidwa ndi nambala yaikulu umasankha ngati polojekiti ndi yotheka.
- Kukana kukusweka pipeline yanu. Kafukufuku wa chitetezo, kuyesa kwa adani, nkhani zopeka zosasidwa, kulemba kovuta — njira yolakwikayi ilibe pano.
Sankhani china pamene
- Mukufuna masomphenya. Mawu okha. Onani njira za masomphenya pamwambapa.
- Mukufuna kulingalira kwakuya kwambiri kwa kutembenuka kumodzi. Pa mavuto ovuta kwambiri a kusanthula a mwayi umodzi, DeepSeek-V4-Pro-0813 · 3BIT-REAP ndicho chida cholemera kwambiri pa 2.2× mtengo wosakanizidwa — pa maitanidwe kumene malire ndicho vuto, ndiko kusinthana koyenera.
- Mukufuna kulondola kwapamwamba kwambiri pa chidziwitso chosowa. Mlingo wa
W8A16— MiMo-V2.5-Pro ndi Hy3 — wafinyidwa mosalimba kwambiri ndipo uli pa pansi pa $0.50 / $2.00, pa maziko aang'ono.
Nambala iliyonse ya maziko pa tsamba ili yasindikizidwa ndi wina m'malo mwathu. Manambala a mtengo ndi mphamvu ndi mtengo wathu weniweni wa katalogi yamoyo, wotsimikizika pogwiritsa ntchito /v1/models ndi kiyi yanu. Onani nokha magwero oyambirira.
Manambala a mayeso a GLM-5.2 (Terminal-Bench 2.1, SWE-bench Pro) ndi onenedwa ndi Z.ai a maziko osafinyidwa; kubwerezanso modziyimira pawokha kukudikirabe. Manambala a Claude Opus 4.8 ndi GPT-5.5 akuwonetsedwa kwa kufananiza kokha. Mtengo wa mndandanda wa API ya GLM-5.2 yeniyeni monga wasindikizidwa ndi Z.ai, Seputembala 2026.
08Mafunso Ofunsidwa Kawirikawiri
Kodi API ya GLM yosasidwa pa Shannon AI ndi chiyani?
Ndi id ya mtundu GLM-5.2-3BIT-REAP pa Shannon AI Gateway: kapangidwe kofinyidwa ka maziko a agent a GLM-5.2 a Z.ai, wotumikiridwa pa gulu lathu la ma GPU popanda wosanjikiza wa kukana ndipo popanda kusefa kwa zotulutsa. Amalipiritsa $0.73 pa ma token miliyoni yolowa ndi $2.34 pa ma token miliyoni yotuluka, ali ndi zenera la context la ma token 262,144, ndipo amafikika pa /v1/chat/completions, /v1/messages ndi /v1/responses.
Chifukwa chiyani GLM 5.2 yosasidwa ndi yotsika mtengo kuposa mitundu ina yaikulu mu katalogi?
Chifukwa cha mmene chafinyidwira. 3BIT-REAP imaphatikiza zolemera za 3-bit ndi REAP, njira yodula akatswiri ya mitundu ya Mixture-of-Experts imene imachotsa akatswiri oyendetsedwa mosowa musanachite quantization. Maziko a MoE ya sparse okhala ndi pafupifupi ma parameter 40B ogwira ntchito pa token, odulidwa ndi kusandulidwa mwanjira imeneyi, amagwiritsa ntchito kukumbukira kochepa kwambiri ndi ma GPU-seconds ochepa kwambiri pa token kuposa mtundu wa dense wa mtundu wofanana, ndipo mtengowo ukusonyeza izi.
Kodi GLM yopanda kukana kumatanthauza kuti mtunduwo ulibe zotchinga zilizonse?
Kumatanthauza kuti Shannon siyiwonjezera kalikonse. Palibe chosankha chinthu cha kukana patsogolo pa mtunduwo ndipo palibe chosefera pa ma token akutuluka mwa iwo. Maziko amasunga khalidwe lililonse limene kuphunzitsidwa kwake kokha kunawapatsa, koma palibe chakunja chimene chimawaposa. Inu ndinu eni wosanjikiza wa ndondomeko, ndipo kugwiritsa ntchito kumayendetsedwa ndi Ndondomeko ya Kagwiritsidwe Kaudindo ya Shannon AI.
Kodi GLM-5.2 3BIT-REAP imathandizira masomphenya kapena zolowetsa za zithunzi?
Ayi. Mtunduwu ndi wa mawu okha. Ngati mukufuna zolowetsa za zithunzi pa endpoint yosasidwa, gwiritsani ntchito imodzi ya mitundu ya masomphenya mu katalogi monga MiniMax-M3, Kimi-K2.6, Kimi-K3, MiMo-V2.5 kapena inkling. GLM-5.2 imathandizira zotulutsa zokonzedwa — zonse ziwiri json_schema ndi response_format — ndi kuyitanitsa zida.
Kodi zenera la context ndi kutalika kwapamwamba kwa zotulutsa ndi zotani?
Ma token 262,144 a context ndi ma token 16,384 a zotulutsa pa pempho lililonse. Kudzaza zenera lonse la context kamodzi kumalipiritsa pafupifupi masenti 19 pa $0.73 pa ma token miliyoni yolowa, ndipo kutulutsa kwa kutalika kwapamwamba kumalipiritsa pansi pa masenti 4.
Kodi uwu ndi mtundu wabwino wa agent wosasidwa wa maloop aatali a kuyitanitsa zida?
Ndicho chimene timagwiritsa ntchito poyamba. Mzera wa GLM waphunzitsidwa mwachindunji pa ntchito za agent kuyambira GLM-4.5, kuyitanitsa zida ndi reasoning_effort zonse zimathandizidwa, ndipo mtengowo ukutanthauza kuti kutembenuka kwa agent zana limodzi kwa ma token makumi awiri chikwi kumalipiritsa pafupifupi madola awiri m'malo mwa khumi ndi limodzi. Kukana ndinso njira yenieni yolephera mu maloop a agent, ndipo palibe wosanjikiza wa kukana pano woti uyambitse kena.
Yesani API ya GLM Yosasidwa
$0.73 kulowa · $2.34 kutuluka · context ya 262K · palibe wosanjikiza wa kukana.
Werengani Zolemba za API Onani Mitundu YonseShannon Lab LLC, New Mexico, USA · kugwiritsa ntchito kumayendetsedwa ndi Ndondomeko ya Kagwiritsidwe Kaudindo
Magwero: Z.ai GLM-5 repository · 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 · NIST CAISI Assessment of Z.ai's GLM-5.2, July 2026. Manambala a mayeso a maziko ndi onenedwa ndi labu ndipo si otsimikizika; amafotokoza GLM-5.2 yosafinyidwa, osati kapangidwe kofinyidwa kameneka. Manambala a mtengo, context ndi mphamvu ndi mtengo wa katalogi ya Shannon AI kuyambira 5 Seputembala 2026.