Asigcini okanye sifikelele kwidatha yomsebenzisi, kwaye asinqumami iakhawunti ngaphandle kokuba igunya elisemthethweni lifune inyathelo lokunyanzelisa.
Isikhokelo seModeli · Kimi-K3 · 3BIT-REAP

I-API ye-Kimi K3 Engacinezelwanga

Inqanaba lezona zakhono ziphezulu kwiSango le-Shannon AI: isiseko seendlela ezininzi seeparamitha ezisithriliyoni ezimbini nesiqingatha esinombono, ukubiza izixhobo nomxholo we-262K — akukho maleko wokwala, kwaye lelona xabiso liphezulu kuluhlu lwethu. Nalu ixesha elo xabiso lifanele ukuhlawulwa ngalo.

Ihlaziywe nge-5 Septemba 2026Isikhokelo seModeliIsiseko: Moonshot AI Kimi K3

Ngokufutshane

I-Kimi-K3-3BIT-REAP yeyona modeli inamandla siyihambisayo yaye ibiza kakhulu: $3.83 kwitoken enye yezigidi ezingenayo, $19.12 kwitoken enye yezigidi eziphumayo. Ine-mbono, umxholo weetoken ezingama-262,144, imveliso ephezulu yeetoken ezingama-16,384, ukubiza izixhobo, imveliso ye-JSON-schema, i-reasoning_effort elungiselelwayo nokusasaza kuzo zontathu iilwimi ze-API — ngaphandle komaleko wokwala nangaphandle kokuhluzwa kwemveliso. Sisigqibo esilungileyo kukubaleka kwe-agentic kwexesha elide, ukuqiqa okunzima kwamanyathelo amaninzi, nomsebenzi wemifanekiso apho iinkcukacha ezicokisekileyo zokubonwayo zibalulekile. Sisigqibo esiphosakeleyo kukwahlula ngeendidi, ukukhupha, incoko, nantoni na oyiqhuba amakhulu amawaka amaxesha ngemini.

Uninzi lwezikhokelo zokukhetha imodeli zibhalwa ngokungathi isakhono simahla. Asisimahla. Umbuzo onomdla ngemodeli yenqanaba eliphezulu asisoze sibe ngu "ingaba ingcono" — idla ngokuba njalo — koko ngu "ingaba ingcono ngokwaneleyo, kulo msebenzi, ukuze kuthethelelwe ityala eliphindwe kahlanu kunemodeli yenqanaba elinye ngezantsi". Esi sikhokelo siyiphendula loo nto ngezibalo zokwenene, kwaye sitsho ngokucacileyo apho ukhetho olutshiphu luphumelela khona.

01Iyintoni ngokwenene i-Kimi K3

I-Kimi K3 yeyona modeli iphambili yobunzima obuvulekileyo ye-Moonshot AI, ekhutshwe ngoJulayi 2026 phantsi kweLayisensi ye-Kimi K3 — isizukulwana sesithathu somnombo owaqala nge-Kimi K2, kwaye luyilo olutsha ngokwenene endaweni yokulungiswa kwemodeli.

2.8T
Iiparamitha zizonke
104B
Ezisebenzayo / itoken
896
Iingcali (ezili-16 ezirawutwayo)
262K
Umxholo kwi-Shannon

Ubume obupapashiweyo yimaleko ezingama-93 — ezingama-69 zisebenzisa i-Kimi Delta Attention, ezingama-24 zisebenzisa ingqwalasela efihlakeleyo yeentloko ezininzi enegeyithi, enye ishinyene — inerawuta ye-Stable LatentMoE ekhetha ezili-16 kwiingcali ezingama-896 kwitoken nganye kunye nezimbini ezabelwanayo. I-Moonshot ixela malunga nokuphucuka okuphindwe kabini nesiqingatha ekusebenzeni kokukhula xa kuthelekiswa ne-Kimi K2 kuloo ndibaniselwano. Umbono ngowemvelo endaweni yokuncanyathiselwa: i-encoder ye-MoonViT-V2 yeeparamitha ezingama-401M ondla isiqu esinye esiphatha umbhalo.

Inani lokusasazeka lifanele ukungeniswa engqondweni: i-K3 ivula iiparamitha ezilikhulu elinesine zezigidigidi kwitoken nganye, xa kuthelekiswa nomda wezingamashumi amathathu ukuya kuma-40 ezigidigidi kuninzi lwekhathalogu yethu. Loo nto iphantse iphindwe kathathu izibalo kwitoken nganye eveliswayo, kwaye sisizathu sonke esenza le modeli ibe phezulu kuluhlu lwamaxabiso. Awuhlawuleli inqanaba lophawu lorhwebo — uhlawulela imizuzwana yeGPU.

Umnombo, ngokufutshane

  • Kimi K2 — iiparamitha ezisithriliyoni esinye, ezingamashumi amathathu anesibini ezigidigidi ezisebenzayo, iingcali ezingama-384 ezinezisibhozo ezirawutwayo kwitoken nganye, yaqeqeshwa kwangaphambili kwiitoken ezisithriliyoni ezilishumi elinesihlanu nesiqingatha isebenzisa i-optimizer ye-MuonClip, yakhelwe ukumisa ukuqhuma kwengqwalasela okonakalisa uqeqesho lwe-MoE kwelo nqanaba.
  • Kimi K2 Thinking — isiseko esinye esafundiswa ukuphithikeza ukuqiqa nokusetyenziswa kwezixhobo, kuxelwa ukuba isimelele ukubiza izixhobo okulandelelanayo okungama-200 ukuya kuma-300 ekubalekeni okunye, ngama-44.9% kwi-Humanity's Last Exam ngezixhobo nangama-60.2% kwi-BrowseComp.
  • Kimi K3 — yakhiwe ngokutsha ngeendlela ezininzi zemvelo nefestile yetoken enye yezigidi ngasentla, ngoqeqesho oluqonda ukucuthwa ukususela kwinqanaba lokulungiswa okubekwe iliso ukuya phambili (ubunzima be-MXFP4, ukusebenza kwe-MXFP8). Yiyo loo nto i-K3 inyamezela ukucinezelwa okunamandla ngcono kuneemodeli ezingazange ziqeqeshwe zikucingile.

Umgca odlulayo ngumsebenzi ozimeleyo wexesha elide: isizukulwana ngasinye saphuculelwa ukuhlala sihambelana kumakhulu amanyathelo endaweni yokuphumelela ubalo-mgangatho lweempendulo ezimfutshane. Sesi sakhono osithengayo.

02Ithetha ntoni i-"3BIT-REAP", kwaye kutheni imodeli ye-2.8T iyidinga

Impumezelo kwisazisi semodeli asilohombiso — ichaza okwenziwe kubunzima phambi kokuba bulayishwe kwiqoqo lethu leeGPU, kwaye sisilumkiso esinyanisekileyo esincanyathiselwe kulo lonke ibango apha. Qala ngezibalo. Imodeli yeeparamitha ezisithriliyoni ezimbini nesiqingatha kwinqanaba lokuchaneka le-16-bit ingama-terabyte amahlanu nesiqingatha obunzima. Kwiibhithi ezintathu kubunzima ngabunye ikufuphi kwi-terabyte enye — isekhulu, kodwa lusasazo olusebenzekayo olunendawo eshiyekileyo ye-KV cache yeetoken ezingama-262K. I-REAP sisiqingatha sesibini: indlela yokuthena iingcali eyila ukuba zeziphi iingcali irawuta eyifikelelayo ngokwenene ize isuse ezikhethwa kanqabe phambi kokucutha endaweni yasemva. Kwimodeli erawuta ezili-16 kuphela kwiingcali ezingama-896 kwitoken nganye leyo ilungele ngokwemvelo — ukusasazeka okunjalo kunomsila omde onegalelo elincinci kakhulu kumsinga wokwenene — kwaye ukuthena kuqala kuthetha ukuba uhlahlo-lwabiwo-mali lweebhithi eziseleyo luchithwa kwiingcali ezibalulekileyo.

Yiba nyanisekile ngomahluko

Ukucinezela kubiza ukuchaneka, yaye akulingani. Kuvela kuqala ekukhunjulweni kweenyani zomsila omde nakumanqaku okugqibela okuchaneka kwemathematika enzima, kancinci kakhulu ekulandeleni imiyalelo, ekumiselweni kokubizwa kwezixhobo, okanye kubume besicwangciso eside se-agentic. Uqeqesho lwe-K3 oluqonda ukucuthwa lulo olwenza esi sakhiwo sibambelele. Kodwa ukuba umsebenzi wakho kukufunyanwa kwenyani engaqhelekanga ngesitshotsho esinye, ungathathi ngokuba imodeli ye-2.8T ecuthiweyo ibetha encinci eyi-8-bit ngokusekelwe kwinani leeparamitha kuphela.

Ikhathalogu yethu isebenzisa iimpumezelo ezintathu njengeleli ngabom. I-3BIT-REAP ihamba kwezona ziseko zinkulu, apho kungekho enye into elinganayo. I-W4A16-AUTOROUND-REAP bubunzima obuyi-4-bit kunye nokusebenza okuyi-16-bit okucuthwe kusetyenziswa i-AutoRound — ukujikeleza okufundiweyo okubuyisa ukuchaneka kokwenene ngaphezu kokujikelezela kweyona ikufuphi ngokungenabuchule — kunye nokuthena okufanayo. I-W8A16 linqanaba elithembeke kakhulu, kwiziseko ezincinci ngokwaneleyo ukuba zingafuni ukucinezelwa okunamandla. Ngokufutshane: impumezelo ikuxelela ukuba kusinde okungakanani kokwasekuqaleni; ixabiso likuxelela ukuba kubiza ntoni ukuyiqhuba.

03Umbono: ukuqonda umfanekiso kukuthengela ntoni

I-Kimi-K3 yenye yeemodeli ezintlanu kwikhathalogu yethu yeemodeli ezilishumi elinambini enombono, yaye yeyona inezakhono kwezo zintlanu. Uninzi lwee-API zombono zorhwebo luqhuba isahluli-ndidi kwimifanekiso elayishiweyo yaye lwala udidi olubanzi lwezingenayo phambi kokuba imodeli ziyibone — into engagqibekanga enengqiqo kubathengi kodwa iyingqinamba enkulu kumsebenzi osemthethweni. Abaphandi bokhuseleko abathatha imifanekiso yesikrini yemveliso yokuxhaphaza, amaqela olinganiselo akha ezawo izahluli-ndidi, umsebenzi wophando lwezomthetho nowophicotho lwedatha, kunye novavanyo lwe-red team konke kubetha kwelo donga. I-Kimi-K3 ayinamaleko wokwala nasihluzi kwimveliso.

Apho umbono we-K3 uzifumanela khona ixabiso lawo kwiimodeli zombono ezitshiphu

  • Imizobo eshinyeneyo neeskimatiki — imizobo yoyilo, iitopoloji zothungelwano, iisekethe: nantoni na apho impendulo ixhomekeke ekulandeleni imigca neeleyibhile endaweni yokubiza izinto ngamagama.
  • Umbhalo omncinci kwimifanekiso yesikrini — imikhondo yemifumba, imveliso yeqonga lolawulo, iiseli zespredishithi. Umsantsa phakathi kwamanqanaba ombono mkhulu kakhulu ekufundekeni kubungakanani obuncinci.
  • Umfanekiso ngaphakathi kombhalo omde — umfanekiso wesikrini ofakwe kumxholo weetoken ezingama-200K, apho imodeli kufuneka inxulumanise into eyibonayo nento eyifunde amawaka angamashumi amahlanu eetoken ngaphambili.
  • Ukuthelekiswa kwemifanekiso emininzi — iinguqulelo ezimbini ze-UI, okanye itshati ithelekiswa netheyibhile eyayivelisayo.

Apho ingazifumaneli khona ixabiso layo: ukubhalwa kwezihloko, umbhalo ongenambhalo, i-OCR yamaxwebhu acocekileyo, kunye nemigca yokusebenza yokufaka iimpawu. Ezo zisonjululwe nge-$0.50 kwitoken enye yezigidi ezingenayo — ukuba umsebenzi wakho wemifanekiso umninzi kwaye into nganye ilula ngokwayo, i-Kimi-K2.6 okanye i-MiniMax-M3 ivelisa iimpendulo ongazukwazi ukuzahlula kwezika-K3.

Imifanekiso itya iitoken kuhlahlo-lwabiwo-mali olunye lweetoken ezingama-262,144 njengombhalo yaye ibhatala kumlinganiselo wezingenayo. Imifanekiso yesikrini embalwa yesisombululo esiphezulu ayisimahla.

04Umsebenzi we-agentic wexesha elide — nto leyo eli nqanaba likhoyo ngenxa yayo

Ukuba kukho umsebenzi omnye othethelela inqanaba eliphezulu, nguwo lo. Imisebenzi emifutshane isilela ngokuxolelekayo: imodeli iyaphosisa, uyaqaphela, uyazama kwakhona, ulahlekelwe yimizuzwana emine. I-agent yemijikelo engama-40 ethatha isebe eliphosakeleyo kumjikelo wesithandathu ichitha imijikelo engama-34 elandelayo yakha ngokucokisekileyo phezu kwempazamo, uze uzokwazi ekugqibeleni sele uyihlawulele yonke itoken ephakathi. Loo kungalingani kokwenza ukuba ixabiso ngetoken lisondele ekubeni ngumlinganiselo ophosakeleyo apha. Into ebalulekileyo lixabiso ngokugqitywa ngempumelelo, elilawulwa kukuba ukubaleka kulahlwa kangaphi. Amanani apapashwe yi-Moonshot ngesiseko se-K3 kubalo-mgangatho lwe-agentic:

BrowseComp91.2
Terminal-Bench 2.188.3
FrontierSWE81.2
DeepSWE67.5
HLE-Full (ngezixhobo)56.0
SWE Marathon42.0

Onke amanani axelwe yi-Moonshot ngesiseko se-Kimi K3 esingacuthwanga, ngoJulayi 2026. Ayingomalinganiselo esakhiwo sethu se-3BIT-REAP kwaye asiwabonisi njengokunjalo.

Isimo sihambelana noko uyilo lwakhelwe kona: i-BrowseComp ne-Terminal-Bench zibalo-mgangatho lwesangqa sezixhobo — imijikelo ephindaphindiweyo yokwenza, ukuqwalasela, ukulungisa. I-SWE Marathon kuma-42.0 sisilinganiso: lolona vavanyo lunzima lokukhowuda kwexesha elide kwiseti, apho nemodeli ephambili isilela kuninzi lwemisebenzi. I-"agentic" ayilulodidi olusonjululweyo; yila ngengqondo yokusilela nokuba ukhetha yiphi imodeli.

Iisetingi ezisebenzayo zokubaleka okude

I-K3 ivula i-reasoning_effort, kwaye apha liqhosha lexabiso lokwenene: iitoken zokuqiqa zibhatala njengeetoken eziphumayo, nge-$19.12 kwitoken enye yezigidi.

  • Ephantsi — ukuthunyelwa kokubizwa kwezixhobo nokukhutshwa okucwangcisiweyo, apho imodeli ikhetha phakathi kokhetho olwaziwayo olumbalwa.
  • Ephezulu — into engagqibekanga yomsebenzi we-agentic wokwenene: ukucwangcisa utshintsho lwamanyathelo amaninzi, ukulungisa iimpazamo ngomkhondo womfumba, ukuvumelanisa imithombo ephikisanayo.
  • Ephezulu kunazo zonke — yigcinele umbizo omnye onzima kakhulu kumgca wokusebenza: isicwangciso esiphezulu kokubaleka, okanye uhlolo lokugqibela. Ukubaleka okuchitha iitoken zokuqiqa ezingama-30K kwephezulu nezingama-90K kwephezulu kunazo zonke kuhlawula i-$1.15 eyongezelelweyo ngenxa yaloo flegi inye.

Yenza uhlahlo-lwabiwo-mali lomxholo ngabom nako: ifestile yethu yiitoken ezingama-262,144 inemveliso ephezulu yeetoken ezingama-16,384, kwaye nge-$3.83 kwitoken enye yezigidi ezingenayo imbali engathenwanga iyanda — umjikelo ngamnye ubhatalisa kwakhona sonke isiqalo.

05Kuthetha ntoni "ukungacinezelwa" kule ndidi

Yonke imodeli kwiSango le-Shannon AI ihanjiswa ngaphandle komaleko wokwala nangaphandle kokuhluzwa kwesiqulatho kwimveliso. Kubaluleke ngakumbi kwinqanaba eliphezulu, kuba imisebenzi abantu abayizisayo kwimodeli yenqanaba eliphezulu yeyona inokukhubekisa isihluzi: uphando lokhuseleko lokuchasa, uhlalutyo lwezixhobo zokuhlasela, ukuphuhliswa kwesahluli-ndidi solinganiselo, ubukrelekrele bezisongelo, intsomi enzima, kunye nalo naluphi na uvavanyo olufuna imodeli ivelise into ozama ukuyibhaqa.

  • Akukho sahluli-ndidi semveliso. Iimpendulo aziskenwa zibhalwe ngokutsha okanye zivalelwe emva kokuveliswa.
  • Akukho sango lezingenayo. Imiyalelo nemifanekiso ayiskrinwa phambi kokuba ifikelele kwimodeli.
  • Ayilomodeli yahlukileyo. Esi sisiseko se-Kimi K3 esihanjiswa ngaphandle komaleko odla ngokuhlala phambi kwaso.
  • Asikokudluliselwa koxanduva. Ukususa umaleko wokwala kudlulisela isigqibo kuwe — bona ipolisi yethu yosetyenziso olunoxanduva.

Kukho inzuzo ethe cwaka kuye nabani na owakha iinkqubo zovavanyo: imodeli engahluzwanga ngumgangatho ozinzileyo. Imodeli eyala ngokungaqikelelekiyo yenza imilinganiselo ibe nengxolo, kuba awunako ukwahlula umda wesakhono kumda womgaqo-nkqubo. Apha ukusilela kukusilela kwesakhono — kuphela kohlobo onokululungisa.

06Ixabiso: $3.83 engenayo, $19.12 ephumayo — nexesha elizifumanela ngalo

Le yeyona modeli ibiza kakhulu kuluhlu lwethu ngomahluko omkhulu: imveliso ixabisa izihlandlo eziyi-9.6 ngaphezu komgangatho wethu we-$2.00 kwaye izihlandlo eziyi-5.2 ngaphezu kwexabiso le-Kimi-K2.6. Nazi izibalo esidlangalaleni.

ImodeliEngenayo / 1MEphumayo / 1MUmbonoJSON schema
Kimi-K3 · 3BIT-REAP$3.83$19.12EweEwe
DeepSeek-V4-Pro-0813 · 3BIT-REAP$1.95$3.90HayiEwe
inkling · W4A16-AUTOROUND-REAP$1.42$6.07EweHayi
Kimi-K2.6 · W4A16-AUTOROUND-REAP$0.78$3.67EweEwe
GLM-5.2 · 3BIT-REAP$0.73$2.34HayiEwe
MiniMax-M3 · 3BIT-REAP$0.50$2.00EweEwe
MiMo-V2.5 · W8A16$0.50$2.00EweEwe

Imiqolo ekhethiweyo, ihlelwe ngexabiso lezingenayo. Zonke iimodeli ezilishumi elinambini zabelana ngomxholo omnye weetoken ezingama-262,144, imveliso ephezulu yeetoken ezingama-16,384, ukusasaza, ukubiza izixhobo kunye ne-reasoning_effort. Iiflegi zesakhono eziphilayo zikwiphepha lamaxwebhu e-API.

Imizekelo emibini esebenzileyo

Owokuqala, ugalelo lweetoken ezingama-120,000 — ukukhutshwa kovimba wekhowudi, ifayilo yolawulo — luvelisa uhlalutyo lweetoken ezingama-3,000. Owesibini, ukubaleka kwe-agent kwemijikelo engama-40: iitoken zezingenayo ezidibeneyo ezimalunga ne-1.2M njengoko umxholo unyuka, iitoken ezingama-90,000 eziphumayo kubandakanya ukuqiqa.

ImekoKimi-K3Kimi-K2.6MiniMax-M3
120K engenayo / 3K ephumayo$0.517$0.105$0.066
Ukubaleka kwe-agent kwemijikelo engama-40$6.32$1.27$0.78

I-K3 ixabisa malunga nezihlandlo eziyi-4.9 zexabiso le-Kimi-K2.6 kumbizo woxwebhu. Iisenti ezingamashumi amahlanu azilutho kanye; zingama-$517 ngemini kwimibizo eliwaka, xa kuthelekiswa nama-$105. Kukubaleka kwe-agent umahluko uma-$5 malunga. Ngoku sebenzisa lona kuphela uvavanyo olubalulekileyo: ukuba umsebenzi ufanele iyure yexesha lomnjineli enge-$100, i-K3 iyazihlawulela kwangoko iphakamisa impumelelo yepasi yokuqala ngamanqaku amahlanu epesenti — kwaye kumsebenzi wexesha elide onzima ngokwenene umsantsa phakathi kwamanqanaba udla ngokuba mkhulu ngakumbi. Ukuba umsebenzi ufanele imizuzwana engama-30 yexesha lakhe nabani na, loo $5 inye ayinakuthethelelwa.

Umthetho esiwusebenzisayo ngokwenene

Nyukela kwi-Kimi-K3 xa ixabiso lokusilela kokubaleka lisiba ngaphezu komahluko wexabiso lokukuqhuba kwimodeli engcono. Loo nto ibeka isigqibo kwixabiso lomsebenzi endaweni yexabiso letoken. Umsebenzi wexabiso eliphezulu, womthamo osezantsi, wexesha elide, nonzima ukuqinisekiswa uya kwi-K3. Umsebenzi womthamo omkhulu, wobungozi obusezantsi, olula ukuzanywa kwakhona awuyi — nokuba inani leeparamitha limangalisa kangakanani.

07I-Kimi-K3 okanye udade otshiphu?

Iimodeli ezimbini zeKimi kwikhathalogu yethu ngamanqwanqwa amabini eleli elinye endaweni yokuba ngabakhuphisani. I-Kimi-K2.6 icuthwe ngamandla amancinci — i-W4A16-AUTOROUND-REAP, ubunzima obuyi-4-bit ngokujikeleza okufundiweyo, xa kuthelekiswa ne-3-bit ye-K3 — kwisiseko esincinci. Inombono, imveliso ye-JSON-schema, ifestile yomxholo efanayo, kwaye ixabisa malunga nesihlanu.

UmsebenziSebenzisaKutheni
I-agent ezimeleyo yemijikelo engama-40Kimi-K3Ukusilela kuyanda; umgangatho wesicwangciso kumjikelo wokuqala ugqiba ukubaleka
Ukwakhiwa ngokutsha kobungakanani bovimba wekhowudiKimi-K3Ifuna umxholo opheleleyo ubanjwe uhambelana, ingekuko nje ufunyanwe kuwo
Imizobo eshinyeneyo, umbhalo omncinci womfanekiso wesikriniKimi-K3Owona msantsa mkhulu wombono phakathi kwamanqanaba kukufundeka nokuqiqa kwendawo
Izilungiso zefayile enye, ezichaziweyo kakuhleKimi-K2.6Umsebenzi omiselweyo, ukuzama kwakhona okutshiphu, akukho mahluko wophahla olinganiselekayo
Ukubhalwa kwezihloko zemifanekiso ngobuninziMiniMax-M3 / MiMo-V2.5Kusonjululwe nge-$0.50 engenayo; imveliso ye-K3 iphindwe izihlandlo eziyi-9.6 ngesihloko esinye
Ukuqiqa okunzima kodwa okungumbhalo kuphelaDeepSeek-V4-ProImveliso ye-$3.90 xa kuthelekiswa ne-$19.12 — yeka i-K3 ukuba awuyifuni imifanekiso

Loo mqolo wokugqibela lolona longo ludla ngokuphoswa. Ukuba umsebenzi wakho awubandakanyi mifanekiso, jonga i-DeepSeek-V4-Pro kuqala. Nayo sisakhiwo se-3BIT-REAP kwisiseko esikhulu, inamandla ekuqiqeni okunzima, kwaye imveliso yayo ixabisa $3.90 xa kuthelekiswa ne-$19.12 ye-K3. Uninzi lomsinga ofikelela kwinqanaba eliphezulu kukuqiqa okungumbhalo kuphela okuhanjiswa kakuhle ngokulinganayo ngumqolo omnye ngezantsi.

08Ukubiza i-API ye-Kimi K3 engacinezelwanga

Imodeli ivezwe njenge-Kimi-K3-3BIT-REAP kuzo zontathu iilwimi, nokusasaza kwinye nganye yaye akukho mahluko wesakhono phakathi kwazo — ngumilo wocingo kuphela: i-/v1/chat/completions (ehambelana ne-OpenAI), i-/v1/messages (ehambelana ne-Anthropic), kunye ne-/v1/responses. Isicelo sombono esinamandla okuqiqa aphakanyisiweyo kumsebenzi onzima wokubonwayo:

curl https://shannon-ai.com/v1/chat/completions \
  -H "Authorization: Bearer $SHANNON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Kimi-K3-3BIT-REAP",
    "reasoning_effort": "high",
    "max_tokens": 4096,
    "stream": true,
    "messages": [{
      "role": "user",
      "content": [
        { "type": "text",
          "text": "Trace the data path in this diagram and list every point where an untrusted input crosses a trust boundary." },
        { "type": "image_url",
          "image_url": { "url": "data:image/png;base64,..." } }
      ]
    }]
  }'

I-K3 ixhasa kokubini i-response_format kunye ne-json_schema epheleleyo, ngoko unako ukubopha iimpikiswano zezixhobo ze-agent okanye unyanzelise ingxelo ingene kwisimo esimisiweyo ngaphandle kokuhlalutya emva koko. Ukubiza izixhobo kokwemvelo yaye yindlela eyayilelweyo yokuqhuba ukubaleka okude. Iiparamitha, ubungqina bokungena, imida yesantya kunye nebala lokudlala eliphilayo zikwiphepha lamaxwebhu e-API.

Izinto ezintathu ezonga imali ngosuku lokuqala

  • Thena isiqalo. Umjikelo ngamnye ubhatalisa kwakhona yonke incoko nge-$3.83 kwitoken enye yezigidi; ukushwankathela imbali yemijikelo engama-40 ibe yinto yobume beetoken ezingama-4K kanye kwonga ngaphezu kwakho nokutshintshwa kweemodeli okuninzi.
  • Nciphisa imifanekiso. Iitoken zemifanekiso ziitoken zezingenayo; umfanekiso wesikrini we-4K olinganiselwe kwinto esafundeka unika impendulo efanayo ngentlawulo encinci.
  • Thumela ngobunzima. Imodeli etshiphu kwimijikelo eqhelekileyo, ne-K3 kwisicwangciso, kwisebe elinzima nakuhlolo lokugqibela, idla ngokuba ngama-80% esakhono ngama-25% etyala.

09Imibuzo ebuzwa rhoqo

Yintoni i-API ye-Kimi K3 engacinezelwanga?

Yimodeli ye-Kimi-K3 3BIT-REAP kwiSango le-Shannon AI: isiseko se-Mixture-of-Experts seendlela ezininzi seeparamitha ezisithriliyoni ezimbini nesiqingatha se-Moonshot AI, esihanjiswa kwiqoqo lethu leeGPU sinomxholo weetoken ezingama-262,144, imveliso ephezulu yeetoken ezingama-16,384, umbono, ukubiza izixhobo, imveliso ecwangcisiweyo ye-JSON-schema kunye nokusasaza, ngaphandle komaleko wokwala nangaphandle kokuhluzwa okusetyenziswa kwimveliso. Sifikeleleka ngo-/v1/chat/completions, ngo-/v1/messages nango-/v1/responses.

Ingaba i-API ye-Kimi K3 engacinezelwanga iyawuxhasa umbono?

Ewe. I-Kimi-K3 yenye yeemodeli ezintlanu kwikhathalogu yethu ezinokuqonda umfanekiso, kwaye yeyona inezakhono eziphezulu. Uthumela imifanekiso ngendlela efanayo obungayithumela ngayo kuyo nayiphi na indidi ehambelana ne-OpenAI, njengeenxalenye zesiqulatho ze-image_url ecaleni kombhalo. Imifanekiso itya umxholo kuhlahlo-lwabiwo-mali olunye lweetoken ezingama-262,144 njengombhalo, kwaye iitoken zemifanekiso zibhatalwa kumlinganiselo wezingenayo we-$3.83 kwitoken enye yezigidi.

Kutheni i-Kimi K3 ixabisa $3.83 engenayo kunye ne-$19.12 ephumayo ngoxa ezinye iimodeli ze-Shannon zizi-$0.50 nezi-$2.00?

Kungenxa yokuba isiseko sineeparamitha ezimalunga nesithriliyoni ezimbini nesiqingatha ezinezilikhulu elinesine zezigidigidi ezisebenzayo kwitoken nganye, xa kuthelekiswa neemodeli ezinezingamashumi amathathu ukuya kuma-40 ezigidigidi ezisebenzayo kwenye indawo kwikhathalogu. Iiparamitha ezisebenzayo ezininzi ngetoken zithetha imizuzwana yeGPU emininzi ngetoken, kwaye ixabiso libonisa ixabiso lokuhambisa endaweni yerhafu yesakhono. Imveliso iphindwe izihlandlo eziyi-9.6 kwixabiso leemodeli zethu zomgangatho, ngoko umbuzo ogqibayo ngowokuba ingaba umlinganiselo ophezulu wempumelelo yepasi yokuqala kumsebenzi omde ufaneleke izihlandlo ezimalunga nezintlanu zetyala na.

Ingaba i-Kimi K3 yimodeli enye ne-Kimi K2 okanye ne-Kimi K2 Thinking?

Hayi. I-Kimi K2 yayiyi-MoE yeeparamitha ezisithriliyoni esinye ezinezingamashumi amathathu anesibini ezigidigidi ezisebenzayo neengcali ezingama-384; i-K2 Thinking yongeza ukuqiqa okuphithikeziweyo namakhonkco amade okubiza izixhobo kwisiseko esinye. I-Kimi K3, ekhutshwe yi-Moonshot AI ngoJulayi 2026, luyilo olwahlukileyo: iiparamitha ezisithriliyoni ezimbini nesiqingatha zizonke, ezilikhulu elinesine zezigidigidi ezisebenzayo, iingcali ezingama-896 ezinezilishumi elinesithandathu ezirawutwayo ngetoken, i-Kimi Delta Attention, kunye ne-encoder yombono yemvelo. I-K2 ne-K3 zabelana ngomnombo, ingekuko ngemodeli.

Ndifanele ukusebenzisa nini i-Kimi-K2.6 endaweni ye-Kimi-K3?

Sebenzisa i-Kimi-K2.6 W4A16-AUTOROUND-REAP xa umsebenzi umiselwe kwaye uza kuwuqhuba amaxesha amaninzi: ukwahlula ngeendidi, ukukhupha, izilungiso zefayile enye, incoko, kunye nemisebenzi yemifanekiso apho ufuna ikakhulu inkcazelo enolwazi endaweni yokuqiqa okucokisekileyo kokubonwayo. Nayo inombono, icuthwe ngamandla amancinci, kwaye ixabisa $0.78 engenayo kunye ne-$3.67 ephumayo, malunga nesihlanu se-K3. Nyukela kwi-Kimi-K3 xa ukubaleka kukude, kunamasebe, okanye kubiza ukuba kuphosakale.

Inemithombo, ayibangwanga nje

Lonke inani loyilo nelobalo-mgangatho apha lipapashwe yi-Moonshot AI okanye libonakala kwikhadi labo lemodeli elivulekileyo, kwaye lichaza isiseko esingacuthwanga endaweni yesakhiwo sethu se-3BIT-REAP. Zijongele ngokwakho.

Zama i-API ye-Kimi K3 engacinezelwanga

Inqanaba lezona zakhono ziphezulu ze-agentic nezombono. Akukho maleko wokwala. Hlawula ngetoken.

Funda amaXwebhu e-API Khangela Zonke iiModeli

Shannon Lab LLC · New Mexico, e-USA · iimodeli zihanjiswa kwiqoqo lethu leeGPU


Iinyani zesiseko nobalo-mgangatho ngamanani apapashwe yi-Moonshot AI nge-Kimi K3 nange-Kimi K2 yaye achaza iimodeli zangasentla ezingacuthwanga; isakhiwo sethu lucutho lwe-REAP lwe-3-bit yaye luya kwahluka. Amaxabiso, umxholo neeflegi zesakhono zezika-Shannon AI, njengoko kunjalo nge-5 Septemba 2026 — amanani aphilayo asoloko engawo abuyiswa yi-/v1/models. Imithombo: Ikhadi lemodeli ye-Kimi K3 · Uvimba we-Kimi K3 · Kimi K2 Thinking.

Zonke iilinki zophando