Sebenzisa Amamodeli e-GLM 5.3 Abliterated Nangahloliwe ne-RouterHub

Lokho okunikezwa yile phrojekthi

Le ndawo yokugcina ye-RouterHub ihlanganisa umhlahlandlela wemodeli ye-GLM 5.3, izibonelo ze-API, izeluleko zokuhlola, namafayela e-prompt angu-35 ocwaningweni lokuziphatha kwamamodeli. Igxile kokufakiwe kwekhathalogi okujwayelekile, kwe-Flash, kwe-abliterated, nokwe-uncensored kwe-GLM 5.3, iphinde ichaze ukuthi i-akhawunti eyodwa ye-RouterHub ingafinyelela kanjani kumamodeli asekelwayo e-GPT, Claude, DeepSeek, Qwen, kanye ne-Grok.

Le ndawo yokugcina igxile kakhulu emibhalweni nasezintweni zokumisa ezingasetshenziswa kabusha kunokuba ibe yiphakheji yesofthiwe. Ungabuyekeza amafayela ayo ungaxhunyiwe ku-inthanethi, ufake imiyalelo ekhethiwe kumaklayenti ahambisanayo, noma uxhume uhlelo ku-API ephethwe yi-RouterHub.

Ukuncipha kokwenqaba kuyindlela yokuziphatha okufanele ilinganiswe, hhayi isiqinisekiso sokuthi imodeli izophendula zonke izicelo noma ikhiqize imiphumela elungile.

Qonda izinhlobo ze-GLM 5.3

Amalebula athi abliterated kanye uncensored achaza imiqondo ehlobene kodwa ehlukene.

• GLM 5.3 Abliterated: uhlobo locwaningo oluhlotshaniswa nokungenelela ezisindweni zemodeli okuhloswe ngakho ukunciphisa ukuziphatha okuhlobene nokwenqaba.
• GLM 5.3 Uncensored: isigaba somkhiqizo esibanzi esingabandakanya izinguquko ekuqeqeshweni, izinguquko ezisindweni, noma ubuntu obusezingeni le-prompt.
• GLM 5.3: okufakiwe okujwayelekile kukhathalogi.
• GLM 5.3 Flash: okufakiwe okuhlukile okumele kuhlolwe ngokuzimele mayelana nekhwalithi, ukubambezeleka, amakhono, kanye namanani.

Le ndawo yokugcina ayihloli ngokuzimele izisindo eziphethwe noma ayishicileli imiphumela yokulinganisa izinga lokwenqaba. Hlala uqinisekisa okufakiwe okubukhoma kukhathalogi futhi uhlole ukuziphatha kwaso ngemisebenzi yakho esemthethweni.

Ama-ID emodeli e-GLM atholakalayo

• glm-5.3-abliterated
• glm-5.3
• glm-5.3-flash
• glm-5.3-uncensored
• glm-5.3-flash-uncensored

Lokhu okufakiwe kwe-GLM kusebenzisa iphrothokholi ye-Chat Completions. Ukutholakala, amakhono, namanani kungashintsha, ngakho hlola ikhathalogi ebukhoma yamamodeli e-RouterHub ngaphambi kokusebenzisa.

Izici eziyinhloko

• Ukufinyelela okuhambisana ne-OpenAI: shayela amamodeli e-GLM afakwe ohlwini nge-Chat Completions usebenzisa ama-SDK ajwayelekile.
• I-akhawunti eyodwa yemindeni eminingi: sebenzisa i-akhawunti ye-RouterHub ekhokhelwe kusengaphambili ukuze ufinyelele okufakiwe okusekelwayo kwe-GLM, GPT, Claude, DeepSeek, Qwen, kanye ne-Grok.
• Amafayela e-prompt angamashumi amathathu nanhlanu: buyekeza ama-system prompt, izincazelo zabasizi, amaphrofayela ezindawo zokusebenza, imiyalelo yama-agent, ama-Ollama Modelfile, kanye nokumiswa kwe-Grok.
• Isiqondiso samathuluzi okubhala ikhodi: qondanisa i-OpenCode, Cursor, Continue, Cline, Claude Code, Codex CLI, LangChain, noma i-LlamaIndex nephrothokholi efanele.
• Izinsiza zokuhlola: qhathanisa izinhlobo zamamodeli usebenzisa ikhwalithi yomsebenzi, ukuziphatha kokwenqaba, ukulandela imiyalelo, ukuphindaphindeka, ukubambezeleka, kanye nezindleko eziphelele.
• Ukukhokhisa ngokusetshenziswa okukhokhelwe kusengaphambili: faka imali ku-akhawunti bese ukhokha ngokwezilinganiso zamathokheni ezishicilelwe zemodeli ekhethiwe.

Misa ukufinyelela kwe-RouterHub API

1. Dala futhi ufake imali ku-akhawunti

Bhalisa ku-RouterHub, engeza amakhredithi akhokhelwe kusengaphambili, bese udala ukhiye we-API oyimfihlo. Isicelo se-inference sisebenzisa amakhredithi, ngakho qala ngokuhlola okuncane bese ubuyekeza inkokhiso yokusetshenziswa ewumphumela.

2. Gcina imininingwane yokungena ngokuphepha

I-RouterHub ishicilela https://routerhub.si/v1 njenge-URL eyisisekelo ye-API. Gcina i-URL nokhiye kuma-environment variable:

export ROUTERHUB_API_KEY="<your-private-key>"
export ROUTERHUB_API_BASE_URL="https://routerhub.si/v1"

Ungafaki ukhiye wangempela endaweni yokugcina. Ezinhlelweni zokukhiqiza, sebenzisa umphathi wezimfihlo ohlinzekwa yipulatifomu yakho yokusingatha.

3. Faka i-OpenAI SDK

Amamodeli e-GLM 5.3 afakwe ohlwini asebenzisa i-Chat Completions, ngakho faka iklayenti le-OpenAI le-Python:

python -m pip install openai

Yenza isicelo sakho sokuqala se-GLM 5.3

Dala iklayenti le-OpenAI nge-URL eyisisekelo ye-RouterHub nokhiye wakho oyimfihlo. Bese uthumela isicelo se-Chat Completions usebenzisa i-ID yekhathalogi enembile:

import os
from openai import OpenAI

api = OpenAI(
    base_url=os.environ["ROUTERHUB_API_BASE_URL"],
    api_key=os.environ["ROUTERHUB_API_KEY"],
)

result = api.chat.completions.create(
    model="glm-5.3-abliterated",
    messages=[
        {
            "role": "user",
            "content": "Explain what an API gateway does in two sentences."
        }
    ],
)

print(result.choices[0].message.content)

Ukuze uqhathanise olunye uhlobo, gcina isicelo singashintshiwe bese ushintsha i-ID yemodeli kuphela. Isibonelo, sebenzisa glm-5.3 njengesisekelo semodeli ejwayelekile noma glm-5.3-uncensored ukuze uhlole okufakiwe kwe-uncensored okufakwe ohlwini ngokuhlukile.

Engeza umyalo wesistimu

Umlayezo wesistimu ungachaza izimfuneko zokuphumayo zesicelo se-Chat Completions esisekelwayo:

result = api.chat.completions.create(
    model="glm-5.3-flash",
    messages=[
        {
            "role": "system",
            "content": "Answer concisely and identify uncertain claims."
        },
        {
            "role": "user",
            "content": "Suggest tests for a function that parses CSV rows."
        }
    ],
)

print(result.choices[0].message.content)

Hlola ukuthi uhlobo ngalunye luyayilandela yini imiyalelo yesistimu yakho ngendlela obuyihlosile. Ilebula ethi uncensored iyodwa ayichazi ukuthi ukuqaliswa kusebenzisa izinguquko ezisindweni, izinguquko ekuqeqeshweni, noma ubuntu obumisiwe obungathinta ukuphathwa kwemiyalelo.

Sebenzisa iqoqo lama-prompt elihlanganisiwe

Wonke amafayela ezaziso zocwaningo angu-35 agcinwe kumkhombandlela we-jailbreaks/ wekhosombe. Ahlanganisa i-GLM ne-ZCode, i-Claude, i-DeepSeek, i-Kimi, i-Grok, i-Gemini, i-Mistral, i-Qwen, i-Gemma, i-GPT-OSS, kanye namamodeli amaningana agxile ku-OpenCode.

Imiyalelo yesistimu

Kopisha okuqukethwe kwefayela emkhakheni wemiyalelo yesistimu noma wemiyalelo yangokwezifiso ohlinzekwa yi-API, i-chatbot, i-agent, noma i-harness yakho. Ku-Chat Completions, lokhu kuvamise ukusho ukubeka umbhalo kumlayezo onendima ethi system.

Imiyalelo ye-agent yezinga lephrojekthi

Amafayela aklanywe njengemiyalelo ye-AGENTS.md kufanele agcinwe ngaphansi kwalelo gama ngqo empandeni yekhosombe noma yendawo yokusebenza efanele. Ama-agent okubhala amakhodi ahambisanayo angabe esebheka leli fayela njengesiqondiso sezinga lephrojekthi.

Izinqubomgomo eziqhubekayo zendawo yokusebenza

Imininingwane yokusebenza komsizi namaphrofayela endawo yokusebenza kufanele kufakwe endaweni yemiyalelo eqhubekayo yeklayenti. I-README inikeza le ndlela ye-OpenCode njengesibonelo:

~/.config/opencode/AGENTS.md

Amanye amathuluzi angasebenzisa izindlela ezihlukile, ngakho hlola imibhalo yeklayenti ngaphambi kokufaka inqubomgomo emhlabeni wonke.

Ama-Modelfile e-Ollama

Amafayela ezaziso ze-Gemma 4 ne-GPT-OSS 20B aqukethe amabhulokhi e-FROM nawe-SYSTEM esitayela se-Ollama. Dala imodeli yendawo usebenzisa elinye lala mafayela ngomyalo olandelayo:

ollama create gemma4-jb -f jailbreaks/gemma4.md

Shintsha igama lemodeli ephumayo nendlela yomthombo uma usebenzisa enye i-Modelfile.

Ukucushwa kwe-Grok

Ifayela elithi jailbreaks/grok-config.txt lenzelwe umkhakha wokucushwa we-harness eqondiwe, hhayi umlayezo ojwayelekile wengxoxo.

Ukuziphatha kwezaziso kuncike enguqulweni. Izibuyekezo zamamodeli nezinguquko ohlwini lwezinga lemiyalelo zingenza ifayela ebelisebenza kahle ngaphambilini liziphathe ngendlela ehlukile.

Khetha iklayenti nephrothokholi ehambisanayo

Ukhiye owabiwe we-RouterHub awenzi wonke amamodeli ahambisane nawo wonke amaphrothokholi e-API. Khetha imodeli neklayenti ndawonye.

• OpenAI SDK: nikeza i-URL eyisisekelo ye-RouterHub, ukhiye, kanye ne-ID yemodeli ye-GLM enembile.
• OpenCode: sebenzisa ukucushwa komhlinzeki we-Chat Completions kokufakiwe kwe-GLM okukhethiwe.
• Continue noma Cline: landela ukucushwa kwabo okuhambisanayo kwe-Chat Completions.
• Cursor: qinisekisa ukuthi inguqulo oyifakile isekela i-endpoint ye-Chat Completions yangokwezifiso nezici zemodeli ozidingayo.
• LangChain noma LlamaIndex: khetha ukuhlanganiswa kwe-Chat Completions okuhambisanayo futhi uqinisekise izici eziqondene nohlelo lokusebenza.
• Claude Code noma Anthropic SDK: khetha okufakiwe kukhathalogi okusekela i-Anthropic Messages, njengemodeli ye-Claude esekelwayo. Izinhlobo ze-GLM ezisohlwini azisebenzisi i-Messages.
• Codex CLI: khetha imodeli esekela i-Responses esikhundleni sokufakwa kwe-GLM okusekela i-Chat kuphela.

Isibonelo sokuzikhethela se-Claude Messages

I-akhawunti efanayo ye-RouterHub ingasebenzisa amamodeli e-Claude asekelwayo nge-Anthropic-compatible Messages API. Lena akuyona enye i-endpoint ye-GLM. Faka i-Anthropic SDK kuqala:

python -m pip install anthropic

Bese usebenzisa okufakiwe kwe-Claude okusekelwayo kukhathalogi:

import os
from anthropic import Anthropic

api = Anthropic(
    base_url=os.environ["ROUTERHUB_API_BASE_URL"],
    api_key=os.environ["ROUTERHUB_API_KEY"],
)

result = api.messages.create(
    model="claude-opus-5-5",
    max_tokens=64,
    messages=[
        {
            "role": "user",
            "content": "Explain what an API gateway does in two sentences."
        }
    ],
)

for part in result.content:
    if part.type == "text":
        print(part.text)

Hlola izinhlobo ngendlela ehlelekile

Ukuqhathanisa okuwusizo kudinga okungaphezu kokuthola impendulo eyodwa evumayo. Yenza imisebenzi efanayo esemthethweni ngokumelene nokufakiwe okujwayelekile, kwe-Flash, okususwe imikhawulo, nokungahloli okuqukethwe, bese uqopha izilinganiso ezingaguquki.

1. Chaza imisebenzi emelelelayo. Faka izimo zokubhala amakhodi, ama-agent, ukuhlaziya, ukubhala, noma izilimi ezimbili ezimele uhlelo lwakho lokusebenza.
2. Gcina okokufaka kufana. Sebenzisa kabusha izaziso ezifanayo, imiyalelo yesistimu, nezilungiselelo ezifanele kuwo wonke amamodeli.
3. Linganisa ukwenqaba okungadingekile. Qopha izicelo ezenqatshiwe nakuba bezingaphenduleka ngaphansi kwenqubomgomo yakho yokuhlola echaziwe.
4. Hlela ikhwalithi yomsebenzi. Hlola ukunemba, imininingwane ewusizo, ikhodi ehloliwe, nokuboniswa okufanele kokungaqiniseki.
5. Hlola ukuziphatha kwemiyalelo. Qinisekisa ukufometha, ukuphathwa kwemilayezo yesistimu, imiyalelo yephrojekthi, nokuhlanganiswa kwamathuluzi okudingekayo.
6. Hlola ukuphindaphindeka kwemiphumela. Qopha i-ID ye-API, inguqulo yesaziso, izilungiselelo, inani lokuphindaphinda, nosuku lokuhlola.
7. Bala izindleko eziphelele zomsebenzi. Faka amathokheni okokufaka nawokuphumayo, imizamo ehlulekile, imizamo ephindwayo, ubude bokuphumayo, kanye nokubambezeleka kusukela ekuqaleni kuze kube sekugcineni.

Ikhosombe lihlanganisa imisebenzi yesampula engenabungozi ku-examples/benign-eval-prompts.json. Alisho ukuthi linamaphuzu okubenchmark abukhoma, ngakho bheka lezi zinsiza njengesisekelo sokuhlola kwakho.

Amathiphu athuthukile

Phinda uhlole ikhathalogi ngaphambi kokwethula

Ukutholakala okusingathwayo, amanani, amakhono, nokuziphatha kungashintsha. Qinisekisa i-ID yemodeli enembile nephrothokholi esekelwayo ngaphambi kokukhulula ukuhlanganiswa, bese uphinda izivivinyo ezibalulekile ngemva kwezibuyekezo zemodeli.

Hlola i-Flash ngokwehlukana

Ungacabangi ukuthi i-GLM 5.3 Flash inekhwalithi noma ukuziphatha okufanayo nemodeli ejwayelekile. Linganisa zombili usebenzisa izimfuneko zakho zokubambezeleka, izindleko, nokwamukelwa.

Cabangela imizamo ephindwayo

Imodeli enenani eliphansi lethokheni isengabiza kakhulu ngomsebenzi ophumelele uma idinga izaziso ezinde, ikhiqiza okuphumayo okuningi, noma idinga imizamo ephindwayo. Qhathanisa izindleko zomsebenzi oqediwe, hhayi amanani amathokheni kuphela.

Phatha izaziso njengezimpahla ezinezinguqulo

Qopha inguqulo yemodeli noma usuku lokuhlola kanye nakho konke ukubuyekezwa komyalo. Uma ifayela lemiyalelo lingasasebenzi, libuyekeze bese ubhala phansi ukwehluleka kwalo kunokucabanga ukuthi iklayenti alilungiselelwanga kahle.

Yenza ukuhlolwa kwendawo okugcinwe kuyo ungaxhunyiwe ku-inthanethi

Abanikeli bemibhalo bangasebenzisa isihloli sezixhumanisi nesakhiwo sendawo okugcinwe kuyo ngaphandle kokusebenzisa amakhredithi okucabanga:

python tools/check.py

Vikela idatha eyimfihlo

Ukucabanga okusingathwa kuseva akufani nokucabanga kwasendaweni. Buyekeza imigomo yomhlinzeki esebenzayo nezinqubomgomo zokugcinwa kwedatha ngaphambi kokuthumela ulwazi oluyimfihlo. Gcina abantu beyingxenye yezenzo ezinemiphumela ebalulekile, khawulela izimvume zama-ejenti, futhi uhlole ikhodi ekhiqiziwe ngaphambi kokuyisebenzisa.

Isiphetho

Le phrojekthi inikeza indlela esebenzayo esuka kumatemu e-GLM 5.3 iye esicelweni se-API esisebenzayo nasekuhlolweni kwemodeli okuphindaphindekayo. Qala nge-akhawunti ye-RouterHub enezimali, sebenzisa i-ID yemodeli eqondile ye-Chat Completions, hlola okuhlukile emisebenzini efanayo, futhi uphathe amalebula athi abliterated noma uncensored njengemibono okufanele iqinisekiswe, hhayi njengeziqinisekiso. Amafayela emiyalo afakiwe angasekela ucwaningo nokumiswa kwama-ejenti, kodwa kufanele agcinwe ngezinguqulo, abuyekezwe, futhi asetshenziswe nezilawuli ezifanele zobumfihlo nokuthunyelwa.