I-API ye-Kimi K3 Engavinjelwe
Izinga eliphezulu lamakhono esangweni le-Shannon AI: isisekelo esinezindlela eziningi samapharamitha angu-2.8T esinombono, ukubiza amathuluzi nomongo we-262K — alukho ungqimba lokwenqaba, futhi yintengo ephakeme kunazo zonke ohlwini lwethu. Nansi isikhathi leyo ntengo efanele ukukhokhelwa ngaso.
Ngamafuphi
U-Kimi-K3-3BIT-REAP uyimodeli enamandla kunazo zonke esizikhonzayo futhi ebiza kunazo zonke: $3.83 ngethokheni eyisigidi engenayo, $19.12 ngeyisigidi ephumayo. Inombono, umongo wamathokheni angu-262,144, okukhishwayo okuphezulu kwamathokheni angu-16,384, ukubiza amathuluzi, okukhishwayo kwe-JSON-schema, i-reasoning_effort elungisekayo kanye nokusakaza kuzo zontathu izilimi ze-API — ngaphandle kongqimba lokwenqaba nangaphandle kokuhlungwa kokukhishwayo. Yisinqumo esifanele ekusebenzeni kwama-ejenti kwebanga elide, ekucabangeni okunzima okunezinyathelo eziningi, nasemsebenzini wezithombe lapho imininingwane emihle yokubona ibalulekile. Yisinqumo esingalungile ekuhlukaniseni ngezigaba, ekukhipheni ulwazi, engxoxweni, nakunoma yini oyiqhuba izinkulungwane eziyikhulu zezikhathi ngosuku.
Imihlahlandlela eminingi yokukhetha imodeli ibhalwe sengathi ikhono limahhala. Alimahhala. Umbuzo othakazelisayo ngemodeli yezinga eliphezulu awulokothi ube ngothi "ingabe ingcono" — ngokuvamile injalo — kodwa uthi "ingabe ingcono ngokwanele, kulo msebenzi, ukuze ifanelekise ibhili ephindwe kahlanu kunemodeli yezinga elilodwa ngezansi." Lo mhlahlandlela uphendula lokho ngezibalo zangempela, futhi usho ngokusobala ukuthi izinketho ezishibhile zinqoba kuphi.
01Iyini ngempela i-Kimi K3
I-Kimi K3 iyimodeli eyinhloko yezisindo ezivulekile ye-Moonshot AI, ekhishwe ngoJulayi 2026 ngaphansi kwe-Kimi K3 License — isizukulwane sesithathu sohlobo olwaqala nge-Kimi K2, futhi kungukwakhiwa okusha ngempela kunokulungiswa kwemodeli ekhona.
Uhlaka olushicilelwe lunezingqimba ezingu-93 — ezingu-69 zisebenzisa i-Kimi Delta Attention, ezingu-24 zisebenzisa i-gated multi-head latent attention, eyodwa iminyene — nge-Stable LatentMoE router ekhetha abangu-16 kubochwepheshe abangu-896 kuthokheni ngalinye kanye nabangu-2 abelanwayo. I-Moonshot ibika cishe ukuthuthuka okuphindwe ka-2.5 ekusebenzeni kahle kokukhula uma kuqhathaniswa ne-Kimi K2 kusuka kuleyo nhlanganisela. Umbono ungowemvelo kunokunanyathiselwa: i-encoder ethi MoonViT-V2 yamapharamitha angu-401M idlulisela kuwo lowo mgogodla ophatha umbhalo.
Inombolo yokusabalala ifanele ukugcinwa engqondweni: i-K3 isebenzisa amapharamitha ayizigidigidi ezingu-104 kuthokheni ngalinye, uqhathanisa nebanga elingu-30 kuya ku-40 billion kweningi lekhathalogu yethu. Lokho cishe kuphindwe kathathu izibalo kuthokheni ngalinye elikhiqizwayo, futhi yisona sizathu sonke sokuthi le modeli iphezulu ohlwini lwentengo. Awukhokheli izinga lebhrendi — ukhokhela amasekhondi ama-GPU.
Uhlobo, ngamafuphi
- Kimi K2 — amapharamitha ayi-1 trillion, angu-32 billion asebenzayo, ochwepheshe abangu-384 abangu-8 abaqondiswa kuthokheni ngalinye, iqeqeshwe ngaphambili ngamathokheni ayi-15.5 trillion isebenzisa i-MuonClip optimizer, yakhelwe ukumisa ukuqhuma kokunaka okucekela phansi ukuqeqeshwa kwe-MoE kulelo zinga.
- Kimi K2 Thinking — isisekelo esifanayo esifundiswe ukuxuba ukucabanga nokusetshenziswa kwamathuluzi, kubikwa ukuthi siyakwazi ukuqhubeka nokubiza amathuluzi okulandelanayo okungu-200 kuya ku-300 ekusebenzeni okukodwa, ngo-44.9% ku-Humanity's Last Exam ngamathuluzi nangu-60.2% ku-BrowseComp.
- Kimi K3 — yakhiwe kabusha inezindlela eziningi zemvelo newindi lamathokheni ayi-1M ngaphesheya, nokuqeqeshwa okucabangela ukukalwa kwezinga kusukela esigabeni sokulungiswa okuqondisiwe kuqhubeke (izisindo ze-MXFP4, ukusebenza kwe-MXFP8). Yingakho i-K3 ikubekezelela kangcono ukucindezelwa okuqinile kunamamodeli angakaze aqeqeshwe ekucatshangelwa lokho.
Umugqa oqhubekayo ngu-ubuejenti bebanga elide: isizukulwane ngasinye senzelwe ukuhlala sihambisana kuwo wonke amakhulu ezinyathelo kunokunqoba ukuhlolwa kwezimpendulo ezimfushane. Yilelo khono olithengayo.
02Kusho ukuthini u-"3BIT-REAP", nokuthi kungani imodeli engu-2.8T ikudinga
Isijobelelo ku-id yemodeli akusona umhlobiso — sisho okwenziwa ezisindweni ngaphambi kokuba zilayishwe eqoqweni lethu lama-GPU, futhi siyisexwayiso esiqotho esinamathiselwe kuzo zonke izimangalo lapha. Qala ngezibalo. Imodeli yamapharamitha angu-2.8 trillion ngokunemba kwamabhithi ayi-16 icishe ibe ngamathebhabhayithi angu-5.6 ezisindo. Ngamabhithi amathathu kusisindo ngasinye iseduze nethebhabhayithi elingu-1.05 — isakhulu, kodwa iyindawo esebenzekayo yokuthumela enendawo esele ye-KV cache yamathokheni angu-262K. I-REAP iyingxenye yesibili: indlela yokuthena ochwepheshe ebheka ukuthi yibaphi ochwepheshe i-router eyifinyelela kubo ngempela bese ikhipha abakhethwa kancane ngaphambi kokukala kunangemuva. Kwimodeli eqondisa abangu-16 kuphela kubochwepheshe abangu-896 kuthokheni ngalinye lokho kuyafanela ngokwemvelo — ukusabalala okungako kunomsila omude ofaka okuncane kakhulu ethrafikhini yangempela — futhi ukuthena kuqala kusho ukuthi isabelo esisele samabhithi sichithwa kubochwepheshe ababalulekile.
Yiba neqiniso ngokushintshana
Ukucindezela kubiza ukunemba, futhi hhayi ngokulinganayo. Kuqala kuvele ekukhumbuleni amaqiniso omsila omude nasemaphuzwini okugcina okunemba kwezibalo ezinzima, kancane kakhulu ekulandeleni imiyalelo, ekwakhiweni kokubiza amathuluzi, noma esakhiweni sohlelo olude lwe-ejenti. Ukuqeqeshwa kwe-K3 okucabangela ukukalwa kwezinga yikho okwenza lokhu kwakhiwa kumelane kahle. Kodwa uma umsebenzi wakho ungukutholwa kwamaqiniso angaziwa ngomzamo owodwa, ungacabangi ukuthi imodeli engu-2.8T ekaliwe idlula encane yamabhithi ayisishiyagalombili ngenani lamapharamitha kuphela.
Ikhathalogu yethu isebenzisa izijobelelo ezintathu njengesitebhisi esihlelekile. U-3BIT-REAP ubekwa ezisekelweni ezinkulu kunazo zonke, lapho kungekho okunye okulinganayo. U-W4A16-AUTOROUND-REAP yizisindo zamabhithi amane nokusebenza kwamabhithi ayi-16 okukalwe kusetshenziswa i-AutoRound — ukwenza inani libe liphelele okufundiwe okubuyisa ukunemba kwangempela kunokwenza inani libe liphelele okuseduze kakhulu — kanye nokuthena okufanayo. U-W8A16 yizinga eliphezulu lokunemba, ezisekelweni ezincane ngokwanele ukuthi zingadingi ukucindezelwa okuqinile. Ngamafuphi: isijobelelo sikutshela ukuthi kungakanani kwangempela okusindile; intengo ikutshela ukuthi kubiza malini ukukuqhuba.
03Umbono: ukuqonda izithombe kukuthengelani
I-Kimi-K3 ingenye yamamodeli amahlanu ekhathalogu yethu yamamodeli ayishumi nambili anombono, futhi yiyo enamakhono kunazo zonke ezinhlanu. Ama-API amaningi ezohwebo ombono aqhuba isihlukanisi ezithombeni ezilayishiwe futhi enqabe isigaba esibanzi salokho okufakwayo ngaphambi kokuba imodeli ikubone — okuyinto ezenzakalelayo enengqondo kubathengi kodwa okuyisithiyo esikhulu emsebenzini osemthethweni. Abacwaningi bezokuphepha abathwebula okukhishwayo kokuxhaphaza, amaqembu okulawula okuqukethwe akha izihlukanisi zawo, umsebenzi wobufakazi bomthetho nowokuhlola idatha, kanye nokuhlolwa kwamaqembu abomvu konke kushayisana nalolo donga. I-Kimi-K3 ayinalo ungqimba lokwenqaba futhi ayinaso isihlungi kokukhishwayo.
Lapho umbono we-K3 uzitholela khona intengo yawo ngaphezu kwamamodeli ombono ashibhile
- Imidwebo eminyene namashadi obunjiniyela — imidwebo yokwakhiwa, ukwakheka kwamanethiwekhi, izifunda zikagesi: noma yini lapho impendulo incike ekulandeleni imigqa namalebula kunokuqamba izinto.
- Umbhalo omncane ezithombeni-skrini — imikhondo yezitaki, okukhishwayo kwetheminali, amaseli espredishithi. Umkhawulo phakathi kwamazinga ombono mkhulu kunazo zonke ekufundekeni ngobukhulu obuncane.
- Isithombe ngaphakathi kombhalo omude — isithombe-skrini esifakwe emongweni wamathokheni angu-200K, lapho imodeli kumele ihlobanise ekubonayo nalokho ekufunde emathokhenini angamashumi amahlanu ezinkulungwane ngaphambili.
- Ukuqhathanisa izithombe eziningi — izinguqulo ezimbili ze-UI, noma ishadi liqhathaniswa netafula elilikhiqizile.
Lapho ungayizuzi khona intengo yawo: ukubhala amagama achazayo, umbhalo ophambene, i-OCR yamadokhumenti ahlanzekile, namapayipi okumaka. Lokho kuxazululiwe ku-$0.50 ngethokheni eyisigidi engenayo — uma umsebenzi wakho wezithombe umningi futhi into ngayinye ilula ngokwayo, i-Kimi-K2.6 noma i-MiniMax-M3 zikhiqiza izimpendulo ongeke uzihlukanise nezakwa-K3.
Izithombe zisebenzisa amathokheni kusuka esabelweni esifanayo samathokheni angu-262,144 njengombhalo futhi zikhokhiswa ngezinga lokungenayo. Izithombe-skrini ezimbalwa ezinesixazululo esiphezulu azimahhala.
04Umsebenzi wama-ejenti webanga elide — okuyilona izinga leli elikhona ngenxa yalo
Uma kukhona umsebenzi owodwa ofanelekisa izinga eliphezulu, yilo. Imisebenzi emifushane yehluleka ngokuxolelekayo: imodeli iyaphosisa, uyabona, uyazama futhi, ulahlekelwe imizuzwana emine. I-ejenti lemizuliswano engamashumi amane elithatha igatsha elingalungile emzuliswaneni wesithupha lichitha imizuliswano engamashumi amathathu nane elandelayo lakha ngokucophelela phezu kwephutha, bese uthola ekugcineni sekukhokhelwe wonke amathokheni aphakathi nendawo. Lokho kungalingani yikho okwenza intengo ngethokheni icishe ibe yisilinganiso esingalungile lapha. Okubalulekile yizindleko ngokuqedwa ngempumelelo ngakunye, ezibuswa ukuthi kukangaki ukusebenza okulahlwa. Amanani ashicilelwe e-Moonshot esisekelo se-K3 ekuhlolweni kwama-ejenti:
Wonke amanani abikwe yi-Moonshot ngesisekelo se-Kimi K3 esingakaliwe, ngoJulayi 2026. Awasiwo ukulinganiswa kokwakhiwa kwethu kwe-3BIT-REAP futhi asiwakhombisi kanjalo.
Isimo sihambisana nalokho ukwakhiwa okwenzelwe khona: i-BrowseComp ne-Terminal-Bench ziyizivivinyo zomjikelezo wamathuluzi — imijikelezo ephindaphindiwe yokwenza, ukubuka, ukubuyekeza. I-SWE Marathon ku-42.0 iyisisindo esiphambene: ukuhlolwa okunzima kunakho konke kokubhala ikhodi kwebanga elide kuleli qoqo, lapho ngisha nemodeli ephambili yehluleka emisebenzini eminingi. U-"we-ejenti" akusona isigaba esesixazululiwe; hlelela ukwehluleka noma yiyiphi imodeli oyikhethayo.
Izilungiselelo ezisebenzayo zokusebenza okude
I-K3 iveza u-reasoning_effort, futhi lapha uyisisombululo sangempela sezindleko: amathokheni okucabanga akhokhiswa njengamathokheni aphumayo, ku-$19.12 ngesigidi.
- Phansi — ukuqondisa ukubiza amathuluzi nokukhipha ulwazi oluhlelekile, lapho imodeli ikhetha phakathi kwezinketho ezimbalwa ezaziwayo.
- Phezulu — okuzenzakalelayo emsebenzini wangempela wama-ejenti: ukuhlela ushintsho olunezinyathelo eziningi, ukususa amaphutha kusuka emkhondweni wesitaki, ukubuyisana kwemithombo ephikisanayo.
- Okuphezulu kunakho konke — kugcinele ukubiza okunzima kunakho konke okukodwa epayipini: uhlelo ekuqaleni kokusebenza, noma ukubuyekezwa kokugcina. Ukusebenza okuchitha amathokheni okucabanga angu-30K ku-phezulu nangu-90K kokuphezulu kunakho konke kukhokha u-$1.15 owengeziwe ngalelo folakhi elilodwa.
Hlela nomongo ngamabomu: iwindi lethu ngamathokheni angu-262,144 nokukhishwayo okuphezulu kwamathokheni angu-16,384, futhi ku-$3.83 ngesigidi esingenayo umlando ongathenwanga uyanqwabelana — umzuliswano ngamunye ukhokhisa kabusha sonke isiqalo.
05Kusho ukuthini u-"okungavinjelwe" kulesi siphetho
Yonke imodeli esangweni le-Shannon AI ikhonzwa ngaphandle kongqimba lokwenqaba nangaphandle kokuhlungwa kokuqukethwe kokukhishwayo. Kubaluleke kakhulu ezingeni eliphezulu, ngoba imisebenzi abantu abayiletha kwimodeli yezinga eliphezulu iyiyona engase icupha isihlungi: ucwaningo lokuphepha oluphikisayo, ukuhlaziywa kwamathuluzi ahlaselayo, ukwakhiwa kwezihlukanisi zokulawula okuqukethwe, ubuhlakani bezinsongo, izinganekwane ezinzima, kanye nanoma yikuphi ukuhlolwa okudinga imodeli ikhiqize into ozama ukuyithola.
- Asikho isihlukanisi sokukhishwayo. Izimpendulo azihlolwa zibhalwe kabusha noma zivinjelwe ngemuva kokukhiqizwa.
- Alikho isango lokungenayo. Imiyalelo nezithombe akuhlolwa ngaphambi kokuba kufike kwimodeli.
- Akusiyo imodeli ehlukile. Yisisekelo se-Kimi K3 esikhonzwa ngaphandle kongqimba oluvame ukuhlala phambi kwaso.
- Akusikho ukudluliswa kwesibopho. Ukususa ungqimba lokwenqaba kudlulisela isahlulelo kuwe — bheka inqubomgomo yethu yokusebenzisa ngokuzibophezela.
Kukhona inzuzo ethulile kunoma ubani owakha izinhlelo zokuhlola: imodeli engahlungiwe iyisisekelo esizinzile. Imodeli enqaba ngendlela engabikezeleki yenza ukulinganisa kube nomsindo, ngoba awukwazi ukwehlukanisa umkhawulo wekhono nomkhawulo wenqubomgomo. Lapha ukwehluleka kungukwehluleka kwekhono — okuwuhlobo olulodwa ongalususa amaphutha kulo.
06Intengo: $3.83 engenayo, $19.12 ephumayo — nokuthi nini iyifanela
Le yimodeli ebiza kunazo zonke ohlwini lwethu ngomkhawulo omkhulu: okukhishwayo kubiza ngo-9.6× phansi lethu lika-$2.00 futhi ngo-5.2× kunalokho i-Kimi-K2.6 ekubizayo. Nazi izibalo obala.
| Imodeli | Engenayo / 1M | Ephumayo / 1M | Umbono | JSON schema |
|---|---|---|---|---|
| Kimi-K3 · 3BIT-REAP | $3.83 | $19.12 | Yebo | Yebo |
| DeepSeek-V4-Pro-0813 · 3BIT-REAP | $1.95 | $3.90 | Cha | Yebo |
| inkling · W4A16-AUTOROUND-REAP | $1.42 | $6.07 | Yebo | Cha |
| Kimi-K2.6 · W4A16-AUTOROUND-REAP | $0.78 | $3.67 | Yebo | Yebo |
| GLM-5.2 · 3BIT-REAP | $0.73 | $2.34 | Cha | Yebo |
| MiniMax-M3 · 3BIT-REAP | $0.50 | $2.00 | Yebo | Yebo |
| MiMo-V2.5 · W8A16 | $0.50 | $2.00 | Yebo | Yebo |
Imigqa ekhethiwe, ihlelwe ngentengo yokungenayo. Wonke amamodeli ayishumi nambili abelana ngomongo ofanayo wamathokheni angu-262,144, okukhishwayo okuphezulu kwamathokheni angu-16,384, ukusakaza, ukubiza amathuluzi kanye no-reasoning_effort. Amafolakhi amakhono aphilayo aku-khasi lemibhalo ye-API.
Izibonelo ezimbili ezisebenziwe
Esokuqala, okungenayo kwamathokheni angu-120,000 — ukuthululwa kwendawo yokugcina ikhodi, ifayela lokulawulwa — okukhiqiza ukuhlaziya kwamathokheni angu-3,000. Esesibili, ukusebenza kwe-ejenti kwemizuliswano engamashumi amane: cishe amathokheni ayi-1.2M angenayo aqoqene njengoba umongo unqwabelana, amathokheni angu-90,000 aphumayo kufaka nokucabanga.
| Isimo | Kimi-K3 | Kimi-K2.6 | MiniMax-M3 |
|---|---|---|---|
| 120K engenayo / 3K ephumayo | $0.517 | $0.105 | $0.066 |
| Ukusebenza kwe-ejenti kwemizuliswano engu-40 | $6.32 | $1.27 | $0.78 |
I-K3 ibiza cishe ngo-4.9× i-Kimi-K2.6 ekubizweni kwedokhumenti. Amasenti angamashumi amahlanu awalutho kanye; angu-$517 ngosuku ekubizweni okuyinkulungwane, uqhathanisa no-$105. Ekusebenzeni kwe-ejenti umehluko ucishe ube ngu-$5. Manje sebenzisa isivivinyo esisodwa esibalulekile: uma umsebenzi ufanele ihora lesikhathi sonjiniyela ngo-$100, i-K3 iyazikhokhela ngomzuzu enyusa ngawo impumelelo yokudlula kokuqala ngamaphuzu ayisihlanu ephesenti — futhi emsebenzini onzima ngempela webanga elide umkhawulo phakathi kwamazinga uvame ukuba mkhulu. Uma umsebenzi ufanele imizuzwana engamashumi amathathu yesikhathi salowo nalowo, lowo $5 ofanayo awuvikeleki.
Umthetho esiwusebenzisa ngempela
Khuphukela ku-Kimi-K3 lapho izindleko zokwehluleka kokusebenza zidlula umehluko wentengo yokukuqhuba kwimodeli engcono. Lokho kubeka isinqumo enanini lomsebenzi kunasentengweni yethokheni. Umsebenzi wenani eliphezulu, wevolumu ephansi, webanga elide, onzima ukuwuqinisekisa uya ku-K3. Umsebenzi wevolumu ephezulu, wobungozi obuphansi, ozanywa kabusha kalula awuyi lapho — kungakhathaliseki ukuthi inani lamapharamitha likhanga kangakanani.
07I-Kimi-K3 noma umfowabo oshibhile?
Amamodeli amabili e-Kimi ekhathalogu yethu ayizitebhisi ezimbili zesikhwelo esisodwa kunokuba yizimbangi. I-Kimi-K2.6 ikalwe kancane — u-W4A16-AUTOROUND-REAP, izisindo zamabhithi amane nokwenza inani libe liphelele okufundiwe, uqhathanisa namabhithi amathathu we-K3 — kwisisekelo esincane. Inombono, okukhishwayo kwe-JSON-schema, iwindi lomongo elifanayo, futhi ibiza cishe ingxenye eyodwa kwezinhlanu.
| Umsebenzi | Sebenzisa | Kungani |
|---|---|---|
| I-ejenti elizimele lemizuliswano engamashumi amane | Kimi-K3 | Ukwehluleka kuyanqwabelana; ikhwalithi yohlelo emzuliswaneni wokuqala inquma ukusebenza |
| Ukwakhiwa kabusha kwezinga lendawo yokugcina ikhodi | Kimi-K3 | Kudinga umongo ogcwele ubanjwe uhambisana, hhayi nje utholwe kuwo |
| Imidwebo eminyene, umbhalo omncane wesithombe-skrini | Kimi-K3 | Umkhawulo wombono omkhulu kunazo zonke phakathi kwamazinga ukufundeka nokucabanga ngendawo |
| Ukuhlela ifayela elilodwa, kucaciswe kahle | Kimi-K2.6 | Umsebenzi okhawuliwe, ukuzama futhi okushibhile, awukho umehluko wophahla olinganisekayo |
| Ukubhalela izithombe amagama achazayo ngevolumu | MiniMax-M3 / MiMo-V2.5 | Kuxazululiwe ku-$0.50 engenayo; okukhishwayo kwe-K3 kuphindwe ka-9.6 ngegama elifanayo |
| Okusinda ngokucabanga kodwa okungumbhalo kuphela | DeepSeek-V4-Pro | $3.90 okuphumayo uqhathanisa no-$19.12 — yeka i-K3 uma ungazidingi izithombe |
Lowo mugqa wokugcina yiwona okonga okuvame ukudlulwa kakhulu. Uma umsebenzi wakho ungabandakanyi izithombe, hlola i-DeepSeek-V4-Pro kuqala. Nayo iwukwakhiwa kwe-3BIT-REAP kwisisekelo esikhulu, iqinile ekucabangeni okunzima, futhi okukhishwayo kwayo kubiza u-$3.90 uqhathanisa no-$19.12 wakwa-K3. Iningi lethrafikhi elifinyelela ezingeni eliphezulu lingukucabanga kombhalo kuphela okukhonzwa kahle ngokufanayo umugqa owodwa ngezansi.
08Ukubiza i-API ye-Kimi K3 engavinjelwe
Imodeli ivezwa njengo-Kimi-K3-3BIT-REAP kuzo zontathu izilimi, ngokusakaza kuzo zonke futhi kungekho mehluko wamakhono phakathi kwazo — kuphela isimo sochungechunge: /v1/chat/completions (ehambisana ne-OpenAI), /v1/messages (ehambisana ne-Anthropic), kanye no-/v1/responses. Isicelo sombono ngomzamo wokucabanga okhushuliwe womsebenzi wokubona onzima:
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 isekela kokubili u-response_format ne-json_schema egcwele, ngakho ungakhawulela izimpikiswano zamathuluzi ze-ejenti noma uphoqe umbiko ube yisimo esimisiwe ngaphandle kokuhlaziya kwangemuva. Ukubiza amathuluzi kungokwemvelo futhi yiyo indlela ehlosiwe yokuqhuba ukusebenza okude. Amapharamitha, ukuqinisekisa, imikhawulo yezinga nendawo yokudlala ephilayo kuse-khasini lemibhalo ye-API.
Izinto ezintathu ezonga imali ngosuku lokuqala
- Thena isiqalo. Umzuliswano ngamunye ukhokhisa kabusha yonke ingxoxo ku-$3.83 ngesigidi; ukufingqa umlando wemizuliswano engu-40 ube yinto yesimo yamathokheni angu-4K kanye konga ngaphezu kokushintshwa kwamamodeli okuningi.
- Nciphisa izithombe. Amathokheni ezithombe angamathokheni angenayo; isithombe-skrini se-4K esishintshwe usayizi saba yinto esafundeka sinikeza impendulo efanayo ngokuncane.
- Qondisa ngobunzima. Imodeli eshibhile emizuliswaneni ejwayelekile, ne-K3 ohlelweni, egatsheni elinzima nasekubuyekezweni kokugcina, kuvame ukuba ngu-80% wekhono ngo-25% webhili.
09Imibuzo ebuzwa njalo
Iyini i-API ye-Kimi K3 engavinjelwe?
Yimodeli i-Kimi-K3 3BIT-REAP esangweni le-Shannon AI: isisekelo se-Mixture-of-Experts esinezindlela eziningi samapharamitha angu-2.8T se-Moonshot AI, esikhonzwa eqoqweni lethu lama-GPU ngomongo wamathokheni angu-262,144, okukhishwayo okuphezulu kwamathokheni angu-16,384, umbono, ukubiza amathuluzi, okukhishwayo okuhlelekile kwe-JSON-schema kanye nokusakaza, ngaphandle kongqimba lokwenqaba nangaphandle kokuhlungwa okusetshenziswa kokukhishwayo. Ifinyeleleka nge-/v1/chat/completions, nge-/v1/messages nange-/v1/responses.
Ingabe i-API ye-Kimi K3 engavinjelwe iyawusekela umbono?
Yebo. I-Kimi-K3 ingenye yamamodeli amahlanu ekhathalogu yethu akwazi ukuqonda izithombe, futhi yiyona enamakhono aphezulu kunazo zonke. Uthumela izithombe ngendlela efanayo obungazithumela ngayo kunoma yisiphi isiphetho esihambisana ne-OpenAI, njengezingxenye zokuqukethwe ze-image_url eceleni kombhalo. Izithombe zisebenzisa umongo kusuka esabelweni esifanayo samathokheni angu-262,144 njengombhalo, futhi amathokheni ezithombe akhokhiswa ngezinga lokungenayo elingu-$3.83 ngesigidi.
Kungani i-Kimi K3 ibiza u-$3.83 engenayo no-$19.12 ephumayo kanti amanye amamodeli e-Shannon angu-$0.50 no-$2.00?
Ngoba isisekelo sicishe sibe namapharamitha angu-2.8T anangu-104B asebenzayo kuthokheni ngalinye, uma kuqhathaniswa namamodeli anamapharamitha asebenzayo angu-30 kuya ku-40 billion kwenye indawo ekhathalogu. Amapharamitha amaningi asebenzayo kuthokheni ngalinye asho amasekhondi ama-GPU amaningi kuthokheni ngalinye, futhi intengo ibonisa izindleko zokukhonza kunentela yamakhono. Okukhishwayo kuphindwe ka-9.6 intengo yamamodeli ethu aphansi, ngakho umbuzo onquma ukuthi ingabe izinga eliphezulu lempumelelo yokuqala emsebenzini omude lifanele cishe izikhathi ezinhlanu zebhili.
Ingabe i-Kimi K3 iyimodeli efanayo ne-Kimi K2 noma ne-Kimi K2 Thinking?
Cha. I-Kimi K2 bekuyi-MoE yamapharamitha ayi-1 trillion enamapharamitha angu-32 billion asebenzayo nochwepheshe abangu-384; i-K2 Thinking yengeza ukucabanga okuxutshiwe nemixhumano emide yokubiza amathuluzi kuleso sisekelo esifanayo. I-Kimi K3, ekhishwe yi-Moonshot AI ngoJulayi 2026, ingukwakhiwa okuhlukile: amapharamitha angu-2.8 trillion esewonke, angu-104 billion asebenzayo, ochwepheshe abangu-896 abangu-16 abaqondiswa kuthokheni ngalinye, i-Kimi Delta Attention, kanye ne-encoder yombono yemvelo. I-K2 ne-K3 zabelana ngohlobo, hhayi ngemodeli.
Ngingayisebenzisa nini i-Kimi-K2.6 esikhundleni se-Kimi-K3?
Sebenzisa i-Kimi-K2.6 W4A16-AUTOROUND-REAP lapho umsebenzi ukhawuliwe futhi ungawuqhuba izikhathi eziningi: ukuhlukanisa ngezigaba, ukukhipha ulwazi, ukuhlela ifayela elilodwa, ingxoxo, kanye nemisebenzi yezithombe lapho udinga khona incazelo enekhono kunokucabanga okuhle kwezokubona. Nayo inombono, ikalwe kancane, futhi ibiza u-$0.78 engenayo no-$3.67 ephumayo, cishe ingxenye eyodwa kwezinhlanu ye-K3. Khuphukela ku-Kimi-K3 lapho ukusebenza kude, kunamagatsha, noma kubiza uma kungahamba kahle.
Sonke isibalo sokwakhiwa nesokuhlolwa lapha sishicilelwe yi-Moonshot AI noma siyabonakala ekhadini labo lemodeli elivulekile, futhi sichaza isisekelo esingakaliwe kunokwakhiwa kwethu kwe-3BIT-REAP. Zihlolele wena.
Zama i-API ye-Kimi K3 engavinjelwe
Izinga eliphezulu lamakhono lama-ejenti nelombono. Alukho ungqimba lokwenqaba. Khokha ngethokheni.
Funda Imibhalo ye-API Phequlula Wonke AmamodeliShannon Lab LLC · New Mexico, USA · amamodeli akhonzwa eqoqweni lethu lama-GPU
Amaqiniso esisekelo nokuhlolwa kungamanani ashicilelwe yi-Moonshot AI e-Kimi K3 nawe-Kimi K2 futhi achaza amamodeli angakaliwe angaphambili; ukwakhiwa kwethu kungukukalwa kwezinga kwe-REAP kwamabhithi amathathu futhi kuzohluka. Amanani, umongo namafolakhi amakhono angawe-Shannon AI, akhona kusukela ngoMandulo 5, 2026 — amanani aphilayo ahlala eyilawo abuyiswa u-/v1/models. Imithombo: Ikhadi lemodeli le-Kimi K3 · Indawo yokugcina ikhodi ye-Kimi K3 · Kimi K2 Thinking.