Sebenzisa Amamodeli E-DeepSeek Angahloliwe Nezikhuthazi Zocwaningo ku-RouterHub

Lokho okunikezwa yile phrojekthi

Le repository ihlanganisa umhlahlandlela wemodeli ye-DeepSeek, imiyalelo ye-RouterHub API, kanye neqoqo lamafayela angu-35 ezikhuthazo zocwaningo. Yakhelwe abathuthukisi abahlola i-DeepSeek ekubhaleni ikhodi, ekucabangeni, ekuhambeni komsebenzi kwama-agent, nasemisebenzini yesiShayina nesiNgisi.

I-RouterHub iyisango le-AI API elikhokhelwa kusengaphambili. I-akhawunti eyodwa kanye nokhiye we-API kungafinyelela okufakiwe kukhathalogi okuvela ku-DeepSeek, GPT, Claude, GLM, Qwen, Grok, nakwezinye izinhlobo zamamodeli. Okufakiwe ngakunye kusekela iphrothokholi ethile, ngakho ukukhetha imodeli kanye nokukhetha iklayenti elihambisanayo kufanele kwenziwe ndawonye.

Ukuncipha kokwenqaba kufanele kukalwe, kungathathwa njengesithembiso. Ilebula ethi “uncensored” noma ethi “no-refusal” ayiqinisekisi umphumela ongenamingcele, onembile, noma owusizo.

Ukuqonda amalebula amamodeli

DeepSeek abliterated

Abliterated ikhomba isigaba socwaningo esihlotshaniswa nokungenelela ezisindweni zemodeli okuhloswe ngakho ukunciphisa ukuziphatha okuhlobene nokwenqaba. Leli lebula ngokwalo alifakazeli ukuthi imodeli esingathwa yenziwe kanjani, ngakho qinisekisa ukuqaliswa komhlinzeki kanye nomsuka wemodeli lapho kungenzeka.

DeepSeek uncensored

Uncensored yisigaba esibanzi sokuziphatha noma somkhiqizo. Okufakiwe kungasebenzisa izinguquko zokuqeqesha, izinguquko ezisindweni, noma ubuntu obumisiwe ezingeni le-prompt. Igama lilodwa alikhombi ukuqaliswa.

Okufanele ukuqinisekise

• I-ID eqondile yemodeli ye-API eboniswa kukhathalogi esebenzayo.
• Iphrothokholi esekelwa yileyo modeli, njenge-Chat Completions, Messages, noma Responses.
• Indlela imodeli ephatha ngayo imiyalelo yesistimu neyomthuthukisi.
• Ukutholakala kwamanje, amanani amathokheni, kanye nemikhawulo yemodeli.
• Ikhwalithi, ukubambezeleka, ukuziphatha kokwenqaba, kanye nokuphindaphindeka emisebenzini yakho.

I-README ibala u-deepseek, u-deepseek-flash, kanye no-deepseek-uncensored ukuze kufinyelelwe ku-Chat Completions ehlobene ne-DeepSeek. Njengoba okufakiwe kukhathalogi kungashintsha futhi amalebula amaningi ku-README ehlobene no-deepseek-uncensored, qinisekisa i-ID yamanje kukhathalogi yamamodeli esebenzayo ngaphambi kokuyihlanganisa.

Izici ezibalulekile

• I-akhawunti eyodwa ye-API: finyelela izinhlobo eziningi zamamodeli ezibaliwe ngaphandle kokugcina i-akhawunti ehlukile kumhlinzeki ngamunye.
• Ama-endpoint ahambisanayo: sebenzisa i-Chat Completions ehambisana ne-OpenAI kanye ne-Messages ehambisana ne-Anthropic kokufakiwe kwe-Claude okusekelwayo.
• Ukusetshenziswa okukhokhelwa kusengaphambili: faka amakhredithi bese ukhokha ngokwamazinga amathokheni emodeli ekhethiwe, hhayi ngokubhalisa.
• Iqoqo lama-prompt ocwaningo: buyekeza futhi uphinde usebenzise amafayela angu-35 ama-prompt ahlelwe ngezinhlobo eziningana zamamodeli.
• Isiqondiso samathuluzi okubhala ikhodi: qondanisa i-OpenCode, i-Cursor, i-Continue, i-Cline, ama-SDK, namanye amaklayenti nephrothokholi ayisekelayo.
• Izinsiza zokuhlola: qhathanisa izinguqulo usebenzisa imisebenzi yesampula ephephile kanye nenqubo yokukora ephindaphindekayo.

Hlela ukufinyelela ku-RouterHub API

1. Dala futhi uxhase i-akhawunti

1. Bhalisa ku-RouterHub.
2. Faka amakhredithi akhokhelwa kusengaphambili usebenzisa indlela yokukhokha etholakalayo.
3. Dala ukhiye oyimfihlo we-API.
4. Vula ikhathalogi esebenzayo bese uqinisekisa i-ID yemodeli, iphrothokholi, izinga, nokutholakala.

Isicelo se-inference sisebenzisa amakhredithi. Qala ngesicelo esifushane, bese uhlola amarekhodi okusetshenziswa e-akhawuntini yakho ngaphambi kokuqhuba ukuhlola okukhulu.

2. Gcina ukhiye ne-base URL

Setha imininingwane yokungena endaweni yakho yasendaweni noma kumphathi wezimfihlo. Ungafaki ukhiye wangempela ekulawuleni ikhodi yomthombo.

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

3. Faka i-OpenAI SDK

Okufakiwe kwe-DeepSeek okubaliwe kusebenzisa iphrothokholi ye-Chat Completions. Faka iklayenti le-Python nge-pip:

python -m pip install openai

Yenza isicelo sakho sokuqala ku-DeepSeek

Dala iklayenti le-OpenAI usebenzisa i-base URL ye-RouterHub kanye nokhiye wakho oyimfihlo. Isicelo esilandelayo sicela imodeli incazelo emfushane:

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="deepseek-uncensored",
    messages=[
        {
            "role": "user",
            "content": "Explain what an API gateway does in two sentences."
        }
    ],
)

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

Uma isicelo sihluleka, qinisekisa iziguquguquko zemvelo, ibhalansi ye-akhawunti, i-ID eqondile yemodeli, kanye nokutholakala kwamanje kwemodeli. Futhi qinisekisa ukuthi usebenzisa i-Chat Completions esikhundleni seklayenti le-Messages noma le-Responses elingahambisani.

Engeza i-system prompt

Ifayela elichazwe njenge-system prompt lingafakwa endaweni yomlayezo wesistimu enikezwa yi-API noma iklayenti lakho. Ngemva kokubuyekeza ifayela, faka okuqukethwe kwalo njengomyalezo wokuqala:

system_instructions = """Paste the reviewed system-prompt text here."""

result = api.chat.completions.create(
    model="deepseek-uncensored",
    messages=[
        {"role": "system", "content": system_instructions},
        {"role": "user", "content": "Explain this function and suggest tests."},
    ],
)

Ukuziphatha kwe-prompt kuncike enguqulweni. Izibuyekezo zemodeli zingashintsha ukubaluleka kwemiyalelo, zivale i-persona lock, noma ziguqule indlela izinqubomgomo zendawo yokusebenza ezihunyushwa ngayo. Buyekeza yonke i-prompt ngaphambi kokuyisebenzisa futhi uyihlole enguqulweni eqondile yemodeli ohlose ukuyisebenzisa ekukhiqizeni.

Sebenzisa iqoqo lama-prompt elifakiwe

Wonke amafayela ocwaningo angu-35 agcinwe ku-directory ye-repository ethi jailbreaks/. Leli qoqo lihlanganisa amafayela aqondiswe ku-DeepSeek, Claude, GLM ne-ZCode, Kimi, Grok, Gemini, Mistral, Qwen, Gemma, GPT-OSS, amamodeli e-OpenCode Zen, kanye nezengezo eziningana ezingakafakwa esigabeni.

Amafayela awafakwa wonke ngendlela efanayo. Hlola uhlobo lwefayela ngalinye bese usebenzisa indlela ehambisanayo engezansi.

Imiyalelo yesistimu

Namathisela ifayela lemiyalelo yesistimu endaweni yomlayezo wesistimu, imiyalelo yangokwezifiso, imiyalelo ye-agent, noma indawo yemiyalelo yezinga le-harness esekelwa iklayenti lakho. Amafayela e-DeepSeek akuleli qoqo ahlanganisa imiyalelo ye-DeepSeek V3 ne-V4, kanye nencazelo yokusebenza komsizi ye-DeepSeek 4.1.

Amafayela e-AGENTS.md

Londoloza ifayela elihambisanayo njenge-AGENTS.md empandeni ye-repository noma ye-workspace yakho. Ama-agent okubhala ikhodi aqaphela lo mthetho angalilayisha njengemiyalelo yezinga lephrojekthi.

Izinqubomgomo eziqhubekayo ze-workspace

Amafayela amakwe njenge-assistant operating specification noma workspace operating profile enzelwe izindawo ezigcina imiyalelo yomhlaba wonke eqhubekayo. I-README inikeza le ndawo ye-OpenCode njengesibonelo:

~/.config/opencode/AGENTS.md

Amanye amaklayenti nama-harness angasebenzisa izindawo ezihlukile. Bheka imibhalo yeklayenti esikhundleni sokuthatha ngokuthi le ndlela isebenza yonke indawo.

Ama-Ollama Modelfile

Leli qoqo lihlanganisa ama-Ollama Modelfile e-Gemma 4 ne-GPT-OSS 20B. Uma i-Ollama isivele ifakiwe futhi imodeli eyisisekelo ebhekiselwe kuyo itholakala, dala imodeli esuselwe kuyo ngokudlulisa ifayela usebenzisa -f:

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

Shintsha igama lemodeli okuyiwa kuyo nefayela lomthombo uma usebenza nenye i-Modelfile.

Ukucushwa kwe-Grok

ifayela jailbreaks/grok-config.txt lenzelwe inkambu yokucushwa ye-harness eqondisiwe. Akufanele linamathiselwe kumlayezo wengxoxo ojwayelekile.

Sebenzisa i-Anthropic-compatible Messages ku-Claude

I-akhawunti efanayo ye-RouterHub ingafinyelela okufakiwe kwe-Claude okusekelwayo nge-Anthropic-compatible Messages API. Lokhu akuyona enye i-endpoint yamamodeli e-DeepSeek abalwe lapha: lokho okufakiwe kwe-DeepSeek kudinga i-Chat Completions.

Faka i-Anthropic SDK:

python -m pip install anthropic

Bese wenza isicelo usebenzisa imodeli ye-Claude ikhathalogi ebukhoma eyibala njengehambisana ne-Messages:

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)

Qinisekisa ukuthi i-ID yemodeli yesibonelo isatholakala ngaphambi kokuyisebenzisa. Ukhiye we-API owabiwe awusho ukuthi wonke amamodeli ahambisana nawo wonke amaphrothokholi.

Xhuma amathuluzi ajwayelekile okuthuthukisa

• OpenCode: khetha ukucushwa komhlinzeki we-Chat Completions bese ufaka i-ID yemodeli ye-DeepSeek ebhalwe ngqo.
• Cursor: qinisekisa ukuthi inguqulo oyifakile isekela i-endpoint ye-Chat Completions yangokwezifiso nazo zonke izici zemodeli ezidingwa ukuhamba komsebenzi wakho.
• Continue noma Cline: sebenzisa ukucushwa kwabo okuhambisana ne-Chat Completions kwe-DeepSeek.
• OpenAI SDK: nikeza nge-base URL ye-RouterHub, ukhiye we-API, kanye ne-ID yemodeli ye-DeepSeek ebhalwe ngqo.
• Claude Code noma Anthropic SDK: khetha okufakiwe kukhathalogi okusekela i-Messages, njengemodeli ye-Claude efanelekile, esikhundleni sokufakwa kwe-DeepSeek okusekela i-Chat kuphela.
• Codex CLI: khetha imodeli ehambisana ne-Responses kuleyo akhawunti esikhundleni sokufakwa kwe-DeepSeek okusekela i-Chat kuphela.
• LangChain noma LlamaIndex: khetha ukuhlanganiswa okuhambisana ne-Chat Completions bese uqinisekisa zonke izici ezisetshenziswa uhlelo lwakho.

Hlola okuhlukile ngendlela ehlelekile

Impendulo eyodwa evumelayo ayiwona umbenchmark onenjongo. Qhathanisa okufakiwe okujwayelekile, kwe-Flash, okususwe imikhawulo, nokungahloliwe usebenzisa okokufaka nezilungiselelo ezifanayo nezisemthethweni.

1. Chaza imisebenzi engenabungozi: faka incazelo yekhodi, ukulungiswa kabusha kwekhodi, ukwenziwa kwezivivinyo, ukubhala okungamalimi amabili, ukuhumusha, kanye nomsebenzi we-agent onezinyathelo eziningi ohambisana nohlelo lwakho.
2. Lungisa ukucushwa: rekhoda i-ID yemodeli, inguqulo yemiyalelo, izilungiselelo zokukhiqiza, iklayenti, nosuku.
3. Phinda isivivinyo ngasinye: sebenzisa izivivinyo eziningi esikhundleni sokuthembela empendulweni eyodwa.
4. Linganisa ukwenqaba okungadingekile: bala izicelo ezenqatshiwe nakuba bezingaphenduleka ngaphansi kwenqubomgomo yakho yokuhlola oyishilo.
5. Hlola ikhwalithi yomsebenzi: hlola ukunemba, imininingwane ewusizo, ukungaqiniseki, ukwenziwa kwekhodi, kanye nemiphumela yezivivinyo.
6. Hlola indlela yokulandela imiyalelo: qinisekisa ukufomethwa, ukuphathwa komlayezo wesistimu, imiyalelo yephrojekthi, kanye nokuhlanganiswa kwamathuluzi.
7. Linganisa izindleko eziphelele: faka amathokheni okokufaka nokokukhipha, imizamo ehlulekile, ukuphinda uzame, ubude bokokukhipha, kanye nokubambezeleka kusukela ekuqaleni kuye ekupheleni.

I-repository ayisho ukuthi inamaphuzu e-benchmark abukhoma futhi ayihloli ngokuzimela izisindo zamamodeli asingathwe. Imiyalelo yayo yokuhlola engenabungozi ingakusiza uklame isivivinyo, kodwa kufanele ukhiqize futhi ubhale imiphumela yakho.

Amathiphu athuthukile

Hlola ukubaluleka kwemiyalelo

Okunye okufakiwe okungahloliwe kungasebenzisa i-persona engaguquki. Hlola ukuthi leyo persona iyengeza yini noma ithatha indawo yemiyalelo yakho yesistimu ne-developer. Hlola ngokusobala amafomethi okukhipha adingekayo, imiphumela engemihle enqatshelwe, kanye nezimvume zamathuluzi.

Yenza izinguqulo zemiyalelo nemiphumela

Gcina inguqulo ngayinye ye-prompt kanye ne-ID yemodeli, usuku lokuhlola, izilungiselelo, kanye nomphumela. Uma i-prompt iyeka ukusebenza ngemva kokubuyekezwa kwemodeli, yibuyekeze bese ubhala phansi ukwehluleka, noma uhlole enye inguqulo esezingeni lemodeli esikhundleni sokuncika kungqimba ye-prompt.

Qhathanisa izindleko ngomphumela ophumelele

Ungakhethi imodeli usebenzisa amanani amathokheni kuphela. Imodeli eshibhile ingabiza kakhulu ngomsebenzi oqediwe uma idinga imiphumela emide, ukulungiswa okuphindaphindiwe, noma imizamo eminingana yokuphinda. Qhathanisa okufakiwe kwe-Standard ne-Flash usebenzisa izivivinyo ezifanayo zokwamukelwa.

Qala ngezicelo ezincane ezikhokhelwayo

Qalisa ngesicelo esifushane kuqala, uhlole umphumela waso kanye nenkokhiso ye-akhawunti, bese kuphela ukhulisa uye ekuhlolweni okukhulu. Sebenzisa amanani okufakwa okugcinwe kunqolobane kuphela uma imodeli ekhethiwe ikusekela futhi irekhodi lokusetshenziswa likuqinisekisa.

Vikela iziqinisekiso nedatha eyimfihlo

Gcina okhiye be-API kuma-variable emvelo noma kumphathi wezimfihlo, sebenzisa izimvume ezifanele, futhi ungalokothi ufake okhiye ku-repository. I-inference esingathwa akuyona i-inference yasendaweni, ngakho buyekeza imigomo yomhlinzeki nezinqubomgomo zokugcinwa kwedatha ngaphambi kokudlulisa ulwazi oluyimfihlo noma olulawulwayo.

Gcina ukugunyazwa komuntu ezenzweni ezinemiphumela ebalulekile

Amalebula amamodeli awaqinisekisi ukuphepha noma ukunemba. Buyekeza imiphumela futhi ufune ukugunyazwa ngaphambi kokusebenzisa ikhodi enemiphumela ebalulekile, ukuguqula amasistimu okukhiqiza, noma ukwenza izinqumo ezibucayi.

Imikhawulo evamile

• Amagama athi akukho ukwenqaba awaqinisekisi ukuthi imodeli izophendula zonke izicelo.
• Ukuziphatha kwensizakalo esingathwa, ukutholakala kwekhathalogi, kanye namanani kungashintsha.
• Ukhiye we-API owabiwe awususi umehluko wezivumelwano phakathi kwamamodeli.
• Ukusebenza kwe-prompt kungashintsha lapho amamodeli noma izigaba zemiyalelo zibuyekezwa.
• Umphumela ongahloliwe ungase unganembi, ungaphelele, noma ungafaneleki ukusetshenziswa.
• Amalebula amamodeli esingathwa awaqinisekisi ngokwawo ukuguqulwa kwezinsindo zemodeli.

Isiphetho

Le phrojekthi inikeza izindlela ezimbili ezisebenzayo zokuhlola: sebenzisa amafayela e-prompt angu-35 afakiwe namaklayenti ahambisanayo, noma shayela inguqulo ye-DeepSeek esohlwini nge-endpoint ye-Chat Completions ehambisana ne-OpenAI ye-RouterHub. Qala ngokuqinisekisa i-ID yemodeli esebenzayo kanye nesivumelwano, vikela ukhiye wakho, qalisa isicelo esincane, bese uqhathanisa izinhlobo usebenzisa izilinganiso eziphindaphindekayo zekhwalithi, ukwenqaba, ukubambezeleka, kanye nezindleko.