We do not hold or access any user's data, and we do not suspend accounts unless a lawful authority requires an enforcement action.
Chitsogozo cha Mtundu · Kimi-K3 · 3BIT-REAP

Kimi K3 API Yosasidwa

Mlingo wapamwamba kwambiri pa Shannon AI Gateway: maziko a multimodal a ma parameter thililiyoni 2.8 okhala ndi masomphenya, kuyitanitsa zida ndi context ya 262K — palibe wosanjikiza wa kukana, ndi mtengo wapamwamba kwambiri pa mndandanda wathu. Nawu nthawi imene mtengowo uyenera kulipiridwa.

Yasinthidwa pa 5 Seputembala 2026Chitsogozo cha MtunduMaziko: Moonshot AI Kimi K3

Chidule

Kimi-K3-3BIT-REAP ndicho mtundu wamphamvu kwambiri umene timatumikira ndipo wokwera mtengo kwambiri: $3.83 pa ma token miliyoni yolowa, $19.12 pa miliyoni yotuluka. Uli ndi masomphenya, context ya ma token 262,144, zotulutsa zapamwamba za ma token 16,384, kuyitanitsa zida, zotulutsa za JSON-schema, reasoning_effort yosinthika ndi streaming pa zilankhulo zonse zitatu za API — popanda wosanjikiza wa kukana ndipo popanda kusefa kwa zotulutsa. Chisankho choyenera pa ntchito za agent za nthawi yayitali, kulingalira kovuta kwa masitepe ambiri, ndi ntchito za zithunzi kumene tsatanetsatane wowoneka ndi wofunika. Chisankho cholakwika pa kuika m'magulu, kutulutsa, chat, ndi chilichonse chimene mumayendetsa nthawi zana chikwi patsiku.

Zitsogozo zambiri za kusankha mtundu zimalembedwa ngati kuti mphamvu ndi yaulere. Sichoncho. Funso losangalatsa lokhudza mtundu wapamwamba kwambiri si "kodi ndi wabwino kuposa" — nthawi zambiri ndi choncho — koma "kodi ndi wabwino mokwanira, pa ntchito iyi, kuti muvomereze bilu yaikulu kasanu kuposa mtundu wa sitepe imodzi pansipo." Chitsogozo ichi chimayankha funsoli ndi masamu enieni, ndipo chimanena momveka kumene zisankho zotsika mtengo zimapambana.

01Chimene Kimi K3 chilidi

Kimi K3 ndi mtundu wapamwamba wa zolemera zapoyera wa Moonshot AI, wotulutsidwa mu Julayi 2026 pansi pa Kimi K3 License — mbadwo wachitatu wa mzera umene unayamba ndi Kimi K2, ndi kapangidwe kenieni katsopano m'malo mwa fine-tune.

2.8T
Ma parameter onse
104B
Ogwira ntchito / token
896
Akatswiri (16 oyendetsedwa)
262K
Context pa Shannon

Kapangidwe kosindikizidwa ndi magawo 93 — 69 ogwiritsa ntchito Kimi Delta Attention, 24 ogwiritsa ntchito gated multi-head latent attention, chimodzi dense — ndi router ya Stable LatentMoE yosankha 16 mwa akatswiri 896 pa token kuphatikiza 2 ogawana. Moonshot imanena kukwera pafupifupi kwa 2.5x mu bwino la scaling kuposa Kimi K2 kuchokera ku kaphatikizidwe kameneko. Masomphenya ndi achibadwa m'malo momangiriridwa pambuyo pake: encoder ya MoonViT-V2 ya ma parameter 401M imadyetsa thunthu lomwelo limene limayendetsa mawu.

Nambala ya sparsity ndi yoyenera kuimvetsa: K3 imagwiritsa ntchito ma parameter 104 biliyoni pa token, poyerekeza ndi 30-mpaka-40-biliyoni la mitundu yambiri mu katalogi yathu. Ndi pafupifupi katatu masamu pa token yopangidwa, ndipo ndicho chifukwa chonse chomwe mtunduwu uli pamwamba pa mndandanda wa mitengo. Simukulipira mlingo wa dzina — mukulipira ma GPU-seconds.

Mzera, mwachidule

  • Kimi K2 — ma parameter thililiyoni 1, 32 biliyoni ogwira ntchito, akatswiri 384 ndi 8 oyendetsedwa pa token, ophunzitsidwa poyamba pa ma token thililiyoni 15.5 pogwiritsa ntchito MuonClip optimizer, womangidwa kuti uyimitse kuphulika kwa attention kumene kumawononga kuphunzitsa kwa MoE pa kukula kumeneko.
  • Kimi K2 Thinking — maziko omwewo ophunzitsidwa kusakaniza kulingalira ndi kugwiritsa ntchito zida, onenedwa kusunga maitanidwe a zida 200 mpaka 300 motsatizana mu kuyenda kumodzi, ndi 44.9% pa Humanity's Last Exam ndi zida ndi 60.2% pa BrowseComp.
  • Kimi K3 — womangidwanso ndi multimodality yachibadwa ndi zenera la ma token 1M pamwambapa, ndi kuphunzitsa kodziwa quantization kuyambira pa gawo la supervised fine-tuning kupitirira (zolemera za MXFP4, zogwira ntchito za MXFP8). Ndicho chifukwa chake K3 imalolera kufinya kolimba bwino kuposa mitundu imene sinaphunzitsidwepo ndi ganizo limeneli.

Mzere waukulu ndi ntchito za agent za nthawi yayitali: mbadwo uliwonse unakonzedwa bwino kuti ukhalebe wogwirizana pakati pa masitepe mazana ambiri m'malo mopambana mayeso a mayankho aafupi. Ndilo mphamvu imene mukugula.

02Chimene "3BIT-REAP" chikutanthauza, ndi chifukwa chake mtundu wa 2.8T umachifuna

Chotsatira pa id ya mtunduwo si chokongoletsera — chimanena chimene chinachitikira zolemera zisanaikidwe pa gulu lathu la ma GPU, ndipo ndicho chenjezo loona lomangiriridwa ku chonena chilichonse pano. Yambani ndi masamu. Mtundu wa ma parameter thililiyoni 2.8 pa kulondola kwa 16-bit ndi pafupifupi terabyte 5.6 za zolemera. Pa ma bit 3 pa kulemera kulikonse ndi pafupi ndi terabyte 1.05 — akadali akulu, koma kutumizidwa kotheka ndi malo otsalira a KV cache ya ma token 262K. REAP ndi theka lachiwiri: njira yodula akatswiri imene imayang'ana akatswiri amene router imawagwiritsadi ntchito ndipo imachotsa amene amasankhidwa mosowa musanachite quantization m'malo mutachita. Pa mtundu umene umayendetsa akatswiri 16 okha mwa 896 pa token, izi ndi zoyenera mwachibadwa — kugawa kotsalira kwa sparse kotere kuli ndi mchira wautali umene umapereka zochepa kwambiri pa magalimoto enieni — ndipo kudula poyamba kumatanthauza kuti bajeti ya bit yotsalira imagwiritsidwa ntchito pa akatswiri ofunika.

Khalani oona pa kusinthana

Kufinya kumalipiritsa kulondola, ndipo mosafanana. Kumaonekera poyamba pa kukumbukira zenizeni za mchira wautali ndi pa mfundo zomaliza za kulondola kwa masamu ovuta, kochepa kwambiri pa kutsatira malangizo, mawonekedwe a maitanidwe a zida, kapena kapangidwe ka ndondomeko yayitali ya agent. Kuphunzitsa kodziwa quantization kwa K3 ndicho chifukwa kapangidwe kameneka kamakhalabe kolimba. Koma ngati ntchito yanu ndi kutenga mfundo yosowa nthawi imodzi, musaganize kuti mtundu wa 2.8T wa quantized umaposa mtundu waung'ono wa 8-bit potengera chiwerengero cha ma parameter chokha.

Katalogi yathu imagwiritsa ntchito zotsatira zitatu ngati makwerero adala. 3BIT-REAP imaikidwa pa maziko akulu kwambiri, kumene palibe china chomwe chingakwane. W4A16-AUTOROUND-REAP ndi zolemera za 4-bit ndi zogwira ntchito za 16-bit za quantized pogwiritsa ntchito AutoRound — kuzungulira kophunzitsidwa kumene kumabwezeretsa kulondola kwenikweni pa kuzungulira kwapafupi kosalongosoka — kuphatikiza kudula komweko. W8A16 ndi mlingo wapamwamba kwambiri wa kulondola, kwa maziko aang'ono okwanira osafunikira kufinya kolimba. Mwachidule: chotsatira chimakuuzani kuchuluka kwa choyambirira chimene chapulumuka; mtengo umakuuzani chimene chikulipiritsa kuyendetsa.

03Masomphenya: chimene kumvetsa zithunzi chikukugulirani

Kimi-K3 ndi imodzi mwa mitundu isanu mu katalogi yathu ya mitundu khumi ndi iwiri yokhala ndi masomphenya, ndipo ndi yamphamvu kwambiri mwa asanuyo. API zambiri za malonda za masomphenya zimayendetsa chosankha chinthu pa zithunzi zotumizidwa ndipo zimakana gulu lalikulu la zolowetsa mtunduwo usanaziwoneko — chizoloŵezi choyenera cha ogula ndi chotchinga chachikulu ku ntchito zoyenera. Ofufuza chitetezo ojambula zithunzi za zotulutsa za exploit, magulu owunika opanga zosankha zinthu zawozawo, ntchito za umboni ndi kuunika ma dataset, ndi kuyeza kwa red-team zonse zimakumana ndi khoma limeneli. Kimi-K3 ilibe wosanjikiza wa kukana ndipo ilibe kusefa pa zotulutsa.

Kumene masomphenya a K3 amayenera mtengo wake kuposa mitundu ya masomphenya yotsika mtengo

  • Ma diagram ndi zithunzi za ukadaulo zodzaza — zithunzi za kapangidwe, mawonekedwe a network, ma circuit: chilichonse chimene yankho limadalira kutsata mizere ndi zizindikiro m'malo motchula zinthu.
  • Malemba aang'ono mu zithunzi za pa skrini — stack traces, zotulutsa za terminal, ma cell a spreadsheet. Kusiyana pakati pa milingo ya masomphenya kumakhala kwakukulu kwambiri pa kuwerengeka pa muyeso waung'ono.
  • Chithunzi mkati mwa malemba aatali — chithunzi cha pa skrini choikidwa mu context ya ma token 200K, kumene mtunduwo uyenera kugwirizanitsa chimene ukuona ndi chimene udawerenga ma token 50,000 m'mbuyomo.
  • Kuyerekeza zithunzi zambiri — mitundu iwiri ya UI, kapena chati poyerekeza ndi tebulo limene linaipanga.

Kumene sichiyenera mtengo wake: kulemba zolongosola, alt text, OCR ya zikalata zoyera, ndi ma pipeline a kuika zizindikiro. Zimenezo zimathetsedwa pa $0.50 pa ma token miliyoni yolowa — ngati ntchito yanu ya zithunzi ndi yayikulu ndipo chinthu chilichonse ndi chosavuta pachokha, Kimi-K2.6 kapena MiniMax-M3 zimapanga mayankho amene simungasiyanitse ndi a K3.

Zithunzi zimagwiritsa ntchito ma token kuchokera ku bajeti yomweyo ya ma token 262,144 monga mawu ndipo zimalipiritsidwa pa mtengo wa kulowa. Zithunzi zochepa za pa skrini zapamwamba za resolution siziulere.

04Ntchito za agent za nthawi yayitali — chimene mlingo uwu ulili

Ngati ntchito imodzi imalungamitsa mlingo wapamwamba kwambiri, ndi iyi. Ntchito zazifupi zimalephera mokhululukika: mtunduwo umalakwitsa, mumazindikira, mumayesanso, mwataya masekondi anayi. Agent wa kutembenuka makumi anayi amene amatenga nthambi yolakwika pa kutembenuka kwachisanu ndi chimodzi amamanga mwatsatanetsatane pa cholakwikacho mu makumi atatu ndi anayi otsatira, ndipo mumazindikira pamapeto pake mutalipira token iliyonse pakati. Kusagwirizana kumeneku ndicho chifukwa mtengo pa token uli pafupi ndi muyeso wolakwika pano. Chimene ndi chofunika ndi mtengo pa kumaliza kopambana, kolamulidwa ndi kawirikawiri komwe kuyenda kumatayidwa. Manambala osindikizidwa a Moonshot a maziko a K3 pa mayeso a agent:

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

Manambala onse ndi onenedwa ndi Moonshot a maziko a Kimi K3 osafinyidwa, Julayi 2026. Sizoyeza za kapangidwe kathu ka 3BIT-REAP ndipo sitikuwaperekera motero.

Mawonekedwewo amagwirizana ndi chimene kapangidwe kanamangidwira: BrowseComp ndi Terminal-Bench ndi mayeso a tool-loop — mizere yobwerezabwereza ya kuchita, kuyang'ana, kukonza. SWE Marathon pa 42.0 ndilo lolemera kwambiri: kuyeza kovuta kwambiri kwa kulemba ma kodi kwa nthawi yayitali mu gululi, kumene ngakhale mtundu wapamwamba kwambiri umalephera ntchito zambiri. "Agentic" si gulu lothetsedwa; konzekerani kulephera mtundu uliwonse umene mwasankha.

Zosinthira zothandiza pa kuyenda kwautali

K3 imasonyeza reasoning_effort, ndipo pano ndi chosinthira chenicheni cha mtengo: ma token a kulingalira amalipiritsidwa ngati ma token a zotulutsa, pa $19.12 pa miliyoni.

  • Low — kuyendetsa maitanidwe a zida ndi kutulutsa kokonzedwa, kumene mtunduwo umasankha pakati pa zosankha zochepa zodziwika.
  • High — chikale cha ntchito zenizeni za agent: kukonzekera kusintha kwa masitepe ambiri, kupeza zolakwika kuchokera ku stack trace, kugwirizanitsa magwero osemphana.
  • Max — sungirani pempho limodzi lovuta kwambiri mu pipeline: ndondomeko pachiyambi cha kuyenda, kapena kuunikanso komaliza. Kuyenda kumene kumagwiritsa ntchito ma token 30K a kulingalira pa high ndi 90K pa max kumalipira $1.15 yowonjezera pa chizindikiro chimodzicho.

Konzekeraninso context mwadala: zenera lathu ndi ma token 262,144 ndi zotulutsa zapamwamba za ma token 16,384, ndipo pa $3.83 pa miliyoni yolowa, mbiri yosadulidwa imawonjezerana — kutembenuka kulikonse kumalipiritsanso prefix yonse.

05Chimene "chosasidwa" chikutanthauza pa endpoint iyi

Mtundu uliwonse pa Shannon AI Gateway umatumikiridwa popanda wosanjikiza wa kukana ndipo popanda kusefa kwa zinthu pa zotulutsa. Ndi zofunika kwambiri pa mlingo wapamwamba, chifukwa ntchito zimene anthu amabweretsa ku mtundu wapamwamba ndizo zoyenerera kwambiri kuyambitsa chosefera: kafukufuku wa chitetezo wa adani, kusanthula kwa zida zowukira, kupanga zosankha zinthu za kuwunika, chidziwitso cha ziwopsezo, nkhani zopeka zovuta, ndi kuyeza kulikonse kumene kumafuna mtunduwo kupanga chinthu chimene mukuyesera kupeza.

  • Palibe chosankha chinthu cha zotulutsa. Mayankho sayang'aniridwa ndikulembedwanso kapena kuletsedwa pambuyo pa kupanga.
  • Palibe chipata cha zolowetsa. Ma prompt ndi zithunzi sizigwidwa musanafike ku mtunduwo.
  • Si mtundu wosiyana. Awa ndi maziko a Kimi K3 otumikiridwa popanda wosanjikiza umene nthawi zambiri umakhala patsogolo pake.
  • Si kusamutsa udindo. Kuchotsa wosanjikiza wa kukana kumasamutsira chiweruzo kwa inu — onani ndondomeko yathu ya kagwiritsidwe kaudindo.

Pali phindu lodekha kwa aliyense amene akumanga machitidwe a kuyeza: mtundu wopanda chosefera ndi maziko okhazikika. Mtundu umene umakana mosayembekezeka umapangitsa muyeso kukhala wosokonezeka, chifukwa simungasiyanitse malire a mphamvu ndi malire a ndondomeko. Pano kulephera ndikolephera kwa mphamvu — ndicho mtundu wokhawo umene mungathe kupeza cholakwika chake.

06Mtengo: $3.83 kulowa, $19.12 kutuluka — ndi pamene umayenera mtengo wakewo

Uwu ndi mtundu wokwera mtengo kwambiri pa mndandanda wathu mosiyana kwambiri: zotulutsa zimalipiritsa 9.6× pansi pathu pa $2.00 ndi 5.2× chimene Kimi-K2.6 imalipiritsa. Nawu masamu poyera.

MtunduKulowa / 1MKutuluka / 1MMasomphenyaJSON schema
Kimi-K3 · 3BIT-REAP$3.83$19.12IndeInde
DeepSeek-V4-Pro-0813 · 3BIT-REAP$1.95$3.90AyiInde
inkling · W4A16-AUTOROUND-REAP$1.42$6.07IndeAyi
Kimi-K2.6 · W4A16-AUTOROUND-REAP$0.78$3.67IndeInde
GLM-5.2 · 3BIT-REAP$0.73$2.34AyiInde
MiniMax-M3 · 3BIT-REAP$0.50$2.00IndeInde
MiMo-V2.5 · W8A16$0.50$2.00IndeInde

Mizere yosankhidwa, yokonzedwa ndi mtengo wa kulowa. Mitundu yonse khumi ndi iwiri imagawana context yomweyo ya ma token 262,144, zotulutsa zapamwamba za ma token 16,384, streaming, kuyitanitsa zida ndi reasoning_effort. Zizindikiro za mphamvu zamoyo zili pa tsamba la zolemba za API.

Zitsanzo ziwiri zogwiritsidwa

Choyamba, zolowetsa za ma token 120,000 — kutulutsa kwa codebase, chikalata cha malamulo — kupanga kusanthula kwa ma token 3,000. Chachiwiri, kuyenda kwa agent kwa kutembenuka makumi anayi: pafupifupi ma token olowa 1.2M ophatikizidwa pamene context ikuwonjezereka, ma token 90,000 otuluka kuphatikizapo kulingalira.

MkhalidweKimi-K3Kimi-K2.6MiniMax-M3
120K kulowa / 3K kutuluka$0.517$0.105$0.066
Kuyenda kwa agent kwa kutembenuka 40$6.32$1.27$0.78

K3 imalipiritsa pafupifupi 4.9× Kimi-K2.6 pa pempho la chikalata. Ndalama makumi asanu ndi zopanda tanthauzo kamodzi; ndi $517 patsiku pa mapempho chikwi, poyerekeza ndi $105. Pa kuyenda kwa agent, kusiyanaku ndi pafupifupi $5. Tsopano gwiritsani ntchito mayeso okhawo ofunika: ngati ntchitoyo ili ndi mtengo wa ola limodzi la nthawi ya injiniya pa $100, K3 imalipira yokha nthawi imene ikweza kupambana kwa kudutsa koyamba ndi mfundo zisanu za pachiwerengero — ndipo pa ntchito zenizeni zovuta za nthawi yayitali, kusiyana pakati pa milingo nthawi zambiri kumakhala kwakukulu kuposa. Ngati ntchitoyo ili ndi mtengo wa masekondi makumi atatu a nthawi ya aliyense, $5 yomweyo sikutetezeka.

Lamulo limene timagwiritsadi ntchito

Kwezerani ku Kimi-K3 pamene mtengo wa kuyenda kulephera umaposa kusiyana kwa mtengo wa kuyendetsa pa mtundu wabwino kwambiri. Izi zimaika chisankho pa mtengo wa ntchitoyo m'malo mwa mtengo wa token. Ntchito ya mtengo waukulu, kuchuluka kochepa, nthawi yayitali, yovuta kutsimikizira zimapita ku K3. Ntchito ya kuchuluka kwakukulu, zoika pang'ono, zoyesanso mosavuta sizipita — ngakhale chiwerengero cha ma parameter chikuchititsa chidwi bwanji.

07Kimi-K3 kapena mbale wotsika mtengo?

Mitundu iwiri ya Kimi mu katalogi yathu ndi makwerero awiri a makwerero amodzi m'malo mopikisana. Kimi-K2.6 idasandulidwa mosalimba kwambiri — W4A16-AUTOROUND-REAP, zolemera za 4-bit ndi kuzungulira kophunzitsidwa, poyerekeza ndi 3-bit ya K3 — pa maziko aang'ono. Ili ndi masomphenya, zotulutsa za JSON-schema, zenera lomwelo la context, ndipo imalipiritsa pafupifupi gawo limodzi mwa magawo asanu a mtengowo.

NtchitoGwiritsani ntchitoChifukwa chake
Agent wodziyimira pawokha wa kutembenuka makumi anayiKimi-K3Kulephera kumawonjezerana; ubwino wa ndondomeko pa kutembenuka koyamba umasankha kuyendayo
Kusintha kwakukulu kwa repositoryKimi-K3Kumafuna context yonse yosungidwa mogwirizana, osati kutengedwako chabe
Ma diagram odzaza, malemba aang'ono a pa skriniKimi-K3Kusiyana kwakukulu kwambiri kwa masomphenya pakati pa milingo ndi kuwerengeka ndi kulingalira kwa malo
Kusintha kwa fayilo imodzi, yofotokozedwa bwinoKimi-K2.6Ntchito yamalire, kuyesanso kotsika mtengo, palibe kusiyana koyezeka kwa malire apamwamba
Kulemba zolongosola za zithunzi zambiriMiniMax-M3 / MiMo-V2.5Zathetsedwa pa $0.50 kulowa; zotulutsa za K3 ndi 9.6× pa chilongosolo chomwecho
Kulingalira kwambiri koma mawu okhaDeepSeek-V4-Pro$3.90 zotulutsa poyerekeza ndi $19.12 — siyani K3 ngati simufuna zithunzi

Mzere womalizawo ndi kusunga ndalama kumene kumaphonyedwa kwambiri. Ngati ntchito yanu sikukhudza zithunzi, onani DeepSeek-V4-Pro poyamba. Ndinso kapangidwe ka 3BIT-REAP pa maziko akulu, wamphamvu pa kulingalira kovuta, ndipo zotulutsa zake zimalipiritsa $3.90 poyerekeza ndi $19.12 ya K3. Magalimoto ambiri amene amafunafuna mlingo wapamwamba ndi kulingalira kwa mawu okha komwe kumatumikiridwa bwino mofanana sitepe imodzi pansipo.

08Kuyitana API ya Kimi K3 Yosasidwa

Mtunduwo umasonyezedwa ngati Kimi-K3-3BIT-REAP pa zilankhulo zonse zitatu, ndi streaming pa iliyonse ndipo palibe kusiyana kwa mphamvu pakati pawo — mawonekedwe a wire okha: /v1/chat/completions (yogwirizana ndi OpenAI), /v1/messages (yogwirizana ndi Anthropic), ndi /v1/responses. Pempho la masomphenya lokhala ndi khama la kulingalira lokwezedwa pa ntchito yovuta ya masomphenya:

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,..." } }
      ]
    }]
  }'

K3 imathandizira zonse ziwiri response_format ndi json_schema yathunthu, choncho mungathe kuletsa maudindo a zida za agent kapena kukakamiza lipoti kukhala mu mawonekedwe okhazikika popanda kufufuza pambuyo pake. Kuyitanitsa zida ndi kwachibadwa ndipo ndicho njira yoyenera yoyendetsera kuyenda kwautali. Ma parameter, chitsimikizo, malire a liwiro ndi malo oyeserera amoyo ali pa tsamba la zolemba za API.

Zinthu zitatu zimene zimasunga ndalama patsiku loyamba

  • Dulani prefix. Kutembenuka kulikonse kumalipiritsanso kukambirana konse pa $3.83 pa miliyoni; kufupikitsa mbiri ya kutembenuka makumi anayi kukhala chinthu cha state cha ma token 4K kamodzi kumasunga zambiri kuposa kusintha mitundu yambiri.
  • Chepetsani zithunzi. Ma token a zithunzi ndi ma token olowa; chithunzi cha pa skrini cha 4K chosinthidwa kukula kukhala chinthu chomwe chikadali choweredwa chimapereka yankho lomwelo pa mtengo wochepa.
  • Tumizani malinga ndi kuvuta. Mtundu wotsika mtengo pa kutembenuka kwamba, ndi K3 pa ndondomeko, nthambi yovuta ndi kuunikanso komaliza, nthawi zambiri ndi 80% ya mphamvu pa 25% ya bilu.

09Mafunso Ofunsidwa Kawirikawiri

Kodi API ya Kimi K3 yosasidwa ndi chiyani?

Ndi mtundu wa Kimi-K3 3BIT-REAP pa Shannon AI Gateway: maziko a Mixture-of-Experts a multimodal a ma parameter thililiyoni 2.8 a Moonshot AI, otumikiridwa pa gulu lathu la ma GPU ndi context ya ma token 262,144, zotulutsa zapamwamba za ma token 16,384, masomphenya, kuyitanitsa zida, zotulutsa zokonzedwa za JSON-schema ndi streaming, popanda wosanjikiza wa kukana ndipo popanda kusefa kogwiritsidwa ntchito pa zotulutsa. Zimafikika kudzera mu /v1/chat/completions, /v1/messages ndi /v1/responses.

Kodi API ya Kimi K3 yosasidwa imathandizira masomphenya?

Inde. Kimi-K3 ndi imodzi mwa mitundu isanu mu katalogi yathu yokhala ndi kumvetsa zithunzi, ndipo ndi yamphamvu kwambiri mwa izo. Mumatumiza zithunzi mofanana ndi mmene mungachitire ku endpoint iliyonse yogwirizana ndi OpenAI, ngati zigawo za content za image_url pamodzi ndi mawu. Zithunzi zimagwiritsa ntchito context kuchokera ku bajeti yomweyo ya ma token 262,144 monga mawu, ndipo ma token a zithunzi amalipiritsidwa pa mtengo wa kulowa wa $3.83 pa miliyoni.

Chifukwa chiyani Kimi K3 imalipiritsa $3.83 kulowa ndi $19.12 kutuluka pamene mitundu ina ya Shannon ndi $0.50 ndi $2.00?

Chifukwa maziko ake ali ndi pafupifupi ma parameter thililiyoni 2.8 okhala ndi 104 biliyoni ogwira ntchito pa token, poyerekeza ndi mitundu ya 30-mpaka-40-biliyoni ogwira ntchito kwina mu katalogi. Ma parameter ambiri ogwira ntchito pa token amatanthauza ma GPU-seconds ambiri pa token, ndipo mtengowo ukusonyeza mtengo wa kutumikira m'malo mwa msonkho wa mphamvu. Zotulutsa ndi 9.6× mtengo wa mitundu yathu yapansi, choncho funso losankha ndikuti kodi mlingo wapamwamba wa kupambana kwa kudutsa koyamba pa ntchito yayitali uli ndi phindu la pafupifupi kasanu kuposa bilu.

Kodi Kimi K3 ndi mtundu womwewo ndi Kimi K2 kapena Kimi K2 Thinking?

Ayi. Kimi K2 inali MoE ya ma parameter thililiyoni 1 yokhala ndi ma parameter 32 biliyoni ogwira ntchito ndi akatswiri 384; K2 Thinking idawonjezera kulingalira kosakanikirana ndi maunyolo aatali a maitanidwe a zida pa maziko omwewo. Kimi K3, yotulutsidwa ndi Moonshot AI mu Julayi 2026, ndi kapangidwe kosiyana: ma parameter thililiyoni 2.8 onse, 104 biliyoni ogwira ntchito, akatswiri 896 ndi 16 oyendetsedwa pa token, Kimi Delta Attention, ndi encoder ya masomphenya yachibadwa. K2 ndi K3 zimagawana mzera, osati mtundu.

Kodi ndiyenera kugwiritsa ntchito Kimi-K2.6 nthawi iti m'malo mwa Kimi-K3?

Gwiritsani ntchito Kimi-K2.6 W4A16-AUTOROUND-REAP pamene ntchitoyo ili ndi malire ndipo mungayendetse kangapo: kuika m'magulu, kutulutsa, kusintha kwa fayilo imodzi, chat, ndi ntchito za zithunzi kumene mumafuna makamaka chifotokozero choyenera m'malo mokulingalira kozama kwa masomphenya. Ilinso ndi masomphenya, yasandulidwa mosalimba, ndipo imalipiritsa $0.78 kulowa ndi $3.67 kutuluka, pafupifupi gawo limodzi mwa magawo asanu a K3. Kwezerani ku Kimi-K3 pamene kuyenda ndi kwautali, kokhala ndi nthambi, kapena kokwera mtengo kulakwitsa.

Zochokera ku magwero, osati zonenedwa chabe

Manambala onse a kapangidwe ndi mayeso pano asindikizidwa ndi Moonshot AI kapena akuwonekera mu kaadi yawo yotseguka ya mtundu, ndipo amafotokoza maziko osafinyidwa m'malo mwa kapangidwe kathu ka 3BIT-REAP. Onani nokha.

Yesani API ya Kimi K3 Yosasidwa

Mlingo wapamwamba kwambiri wa agent ndi masomphenya. Palibe wosanjikiza wa kukana. Lipirani pa token.

Werengani Zolemba za API Onani Mitundu Yonse

Shannon Lab LLC · New Mexico, USA · mitundu yotumikiridwa pa gulu lathu la ma GPU


Zoona za maziko ndi mayeso ndi manambala osindikizidwa a Moonshot AI a Kimi K3 ndi Kimi K2 ndipo amafotokoza mitundu ya upstream yosafinyidwa; kapangidwe kathu ndi quantization ya 3-bit REAP ndipo kadzasiyana. Mitengo, context ndi zizindikiro za mphamvu ndi za Shannon AI, zapano kuyambira 5 Seputembala 2026 — manambala amoyo ndi omwe amabwezedwa ndi /v1/models nthawi zonse. Magwero: Kaadi ya mtundu ya Kimi K3 · Repository ya Kimi K3 · Kimi K2 Thinking.

All research links