Hatisi kuchengeta kana kuwana data remushandisi chero upi zvake, uye hatimisi maaccount kunze kwekunge authority iri pamutemo yatida kuti tidaro.
Gwaro Remodhi · Kimi-K3 · 3BIT-REAP

API yeKimi K3 Isina Kudzorwa

Mhando yepamusoro paGedhi reShannon AI: hwaro hwemapharamita matiriyoni 2.8 hunoshandisa zvakawanda hune kuona, kudana zvishandiso necontext ye262K — hapana chidzitiro chekuramba, uye ndiwo mutengo mukuru kupfuura yose pane rondedzero yedu. Heino nguva inokodzera kubhadhara mutengo iwoyo.

Rakavandudzwa musi wa5 Gunyana 2026Gwaro RemodhiHwaro: Moonshot AI Kimi K3

Pfupiso

Kimi-K3-3BIT-REAP ndiyo modhi yakasimba kupfuura dzose yatinoshandisa uye inodhura kupfuura dzose: $3.83 pamatokeni miriyoni anopinda, $19.12 pamiriyoni anobuda. Ine kuona, context yematokeni 262,144, zvinobuda zvakanyanya zvematokeni 16,384, kudana zvishandiso, zvinobuda zveJSON-schema, reasoning_effort inogadziriswa nekuyerera pamitauro yose mitatu yeAPI — pasina chidzitiro chekuramba uye pasina sefa pane zvinobuda. Ndiyo sarudzo yakanaka pakumhanya kwemaagent kwenguva refu, kufungisisa kwakaoma kwematanho akawanda, uye basa remifananidzo apo zvidiki zvinoonekwa zvinokosha. Sarudzo isiriyo pakurongedza, kubvisa ruzivo, kutaurirana, nechero chinhu chaunomhanyisa zviuru zana pazuva.

Magwaro akawanda ekusarudza modhi anonyorwa sekuti unyanzvi hwakasununguka. Hausi. Mubvunzo unonakidza pamusoro pemodhi yepamusoro hausi kuti "iri nani here" — kazhinji iri nani — asi ndewekuti "iri nani zvakakwana here, pabasa iri, kuti bhiri rinopetwa kashanu pane modhi iri chidanho chimwe pasi rikakodzera." Gwaro rino rinopindura izvozvo nemasvomhu chaiwo, uye rinotaura pachena kuti sarudzo dzakachipa dzinokunda kupi.

01Kuti Kimi K3 chii chaizvo

Kimi K3 ndiyo modhi huru yeMoonshot AI ine huremu hwakavhurika, yakaburitswa muna Chikunguru 2026 pasi peKimi K3 License — chizvarwa chechitatu chedzinza rakatanga neKimi K2, uye magadzirirwo matsva chaiwo pane kugadziriswa kwemodhi yekare.

2.8T
Mapharamita ose
104B
Anoshanda / tokeni
896
Nyanzvi (16 dzinotumirwa)
262K
Context paShannon

Chimiro chakabudiswa ndechemadanho 93 — 69 achishandisa Kimi Delta Attention, 24 achishandisa gated multi-head latent attention, rimwe rakazara — nerouter yeStable LatentMoE inosarudza nyanzvi 16 kubva pa896 patokeni imwe neimwe pamwe ne2 dzinogovaniswa. Moonshot inoshuma kuvandudzwa kunopetwa ka2.5 kwekushanda kwekukura kuchienzaniswa neKimi K2 kubva pamusanganiswa iwoyo. Kuona ndekwepachayo pane kwakanamirwa: muisi weMoonViT-V2 une mapharamita 401M unopa danda rimwe chetero rinobata mavara.

Nhamba yekushomeka yakakodzera kunzwisiswa: K3 inoshandisa mapharamita mabhiriyoni 104 patokeni imwe neimwe, kuenzaniswa nemabhiriyoni makumi matatu kusvika makumi mana pakatarogu yedu yakawanda. Ndiwo masvomhu anopetwa katatu patokeni inogadzirwa, uye ndicho chikonzero chose chekuti modhi iyi itungamirire rondedzero yemitengo. Hausi kubhadhara mhando yezita — uri kubhadhara masekondi eGPU.

Dzinza, muchidimbu

  • Kimi K2 — mapharamita etiriyoni imwe, mabhiriyoni 32 anoshanda, nyanzvi 384 dzine 8 dzinotumirwa patokeni imwe neimwe, yakadzidziswa pamatokeni matiriyoni 15.5 ichishandisa MuonClip optimizer, yakavakirwa kumisa kuputika kwekutarisa kunoparadza kudzidziswa kweMoE pachiyero ichocho.
  • Kimi K2 Thinking — hwaro humwe chetehwo hwakadzidziswa kupindirana kufungisisa nekushandisa zvishandiso, yakashumwa ichichengeta kudana zvishandiso zvakateedzana 200 kusvika 300 mukumhanya kumwe chete, ine 44.9% paHumanity's Last Exam ine zvishandiso ne60.2% paBrowseComp.
  • Kimi K3 — yakavakwazve nekushandisa zvakawanda kwepachayo nehwindo rematokeni 1M kunevaridzi, ine kudzidziswa kunoziva quantization kubva padanho rekugadzirisa rakatariswa zvichienda mberi (huremu hweMXFP4, maactivation eMXFP8). Ndicho chikonzero K3 ichitakura kudzikisirwa kwakaomarara zviri nani pane mamodhi asina kumbodzidziswa nazvo mupfungwa.

Nzira inotevedzwa ndeyokushanda kwakazvimirira kwenguva refu: chizvarwa chega chega chakagadzirirwa kuramba chichinzwisisika pamatanho mazana pane kukunda pamiyero yemhinduro pfupi. Ndihwo unyanzvi hwauri kutenga.

02Kuti "3BIT-REAP" inorevei, uye sei modhi ye2.8T ichizvida

Chipfupiso chiri pazita remodhi hachisi chishongo — chinotaura zvakaitwa kuhuremu husati hwaiswa pacluster yedu yemaGPU, uye ndicho chirevo chechokwadi chakanamirwa pachirevo chega chega chiri pano. Tanga nemasvomhu. Modhi ine mapharamita matiriyoni 2.8 pachiyero che16-bit inenge materabhaiti 5.6 ehuremu. Pamabhiti matatu pahuremu humwe nehumwe inosvika pedyo neterabhaiti 1.05 — ichiri hombe, asi kuiisa panzvimbo kunobvira uine nzvimbo yakasara yeKV cache yematokeni 262K. REAP ndicho chikamu chechipiri: nzira yekubvisa nyanzvi inoongorora kuti ndedzipi nyanzvi dzinonyatsosvikirwa nerouter yobvisa dzisingawanzosarudzwa isati yaita quantization pane mushure mayo. Pamodhi inotumira nyanzvi 16 chete kubva pa896 patokeni imwe neimwe ndiko kukodzera chaiko — kugoverwa kwakashoma kudaro kune muswe murefu usingabatsiri zvachose pazvinhu zvinoitika chaizvo — uye kubvisa kutanga kunoreva kuti mabhiti asara anoshandiswa panyanzvi dzinokosha.

Iva wechokwadi pamusoro pekusiyana

Kudzikisira kunodhura kururama, uye kwete zvakaenzana. Kunotanga kuonekwa pakuyeuka zvokwadi zvemuswe murefu uye pamapoinzi ekupedzisira ekururama kwemasvomhu akaoma, kushoma zvikuru pakutevedzera mirairo, chimiro chekudana zvishandiso, kana chimiro chehurongwa hurefu hweagent. Kudzidziswa kwaK3 kunoziva quantization ndicho chikonzero chivakwa ichi chichisimba. Asi kana basa rako riri rekutsvaga zvokwadi zvisingazivikanwi kamwe chete, usafunga kuti modhi ye2.8T yakadzikiswa inokunda diki ye8-bit nenhamba yemapharamita chete.

Katarogu yedu inoshandisa zvipfupiso zvitatu semanera akarongwa nemaune. 3BIT-REAP inoiswa pahwaro hukuru kupfuura hwose, apo hapana chimwe chinokwana. W4A16-AUTOROUND-REAP ihuremu hwe4-bit nemaactivation e16-bit akaitwa quantization neAutoRound — kutenderedza kwakadzidzwa kunodzorera kururama chaiko pane kutenderedza kwakapusa kune iri pedyo — pamwe nekubvisa nyanzvi kumwe chetekwo. W8A16 ndiyo mhando ine kutendeseka kwakanyanya, yehwaro hudiki husingadi kudzikisirwa kwakaomarara. Muchidimbu: chipfupiso chinokuudza kuti chikamu chakadii chezvekutanga chakapona; mutengo unokuudza kuti kuishandisa kunodhura marii.

03Kuona: chaunotenga nekunzwisisa mifananidzo

Kimi-K3 ndeimwe yemamodhi mashanu mukatarogu yedu yemamodhi gumi nemaviri ane kuona, uye ndiyo ine unyanzvi hwakanyanya pamashanu. MaAPI mazhinji ekutengeserana anoona anomhanyisa muongorori pamifananidzo yakaiswa uye anoramba boka guru rezvinopinda modhi isati yatomboiona — chinhu chine musoro kune vatengi asi chipingamupinyi chikuru pabasa riri pamutemo. Vaongorori vezvekuchengetedza vachitora masikirinishoti ezvinobuda, mapoka ekudzora zvirimo achivaka vaongorori vawo, basa reforensic nekuongorora dhata, nekuongorora kwemapoka anopikisa zvose zvinosangana nedonhodzi iroro. Kimi-K3 haina chidzitiro chekuramba uye haina sefa pane zvinobuda.

Apo kuona kwaK3 kunokodzera mutengo wako kupfuura mamodhi anoona akachipa

  • Madhayagiramu akaunganidzana nemapurani — madhayagiramu emagadzirirwo, zvimiro zvemanetiweki, masekiti: chero chinhu apo mhinduro inotevera mitsetse nemarebhuru pane kudana zvinhu nemazita.
  • Mavara maduku mumasikirinishoti — stack traces, zvinobuda pamutsara wemirairo, masero espredishiti. Mukaha uri pakati pemhando dzekuona ndiwo mukuru pakuverengeka pachiyero chiduku.
  • Mufananidzo mukati memavara marefu — sikirinishoti yakaiswa mucontext yematokeni 200K, apo modhi inofanira kubatanidza zvainoona nezvayakaverenga matokeni zviuru makumi mashanu shure.
  • Kuenzanisa mifananidzo mizhinji — vhezheni mbiri dzeUI, kana chati ichienzaniswa netafura yakaigadzira.

Apo isingakodzeri mutengo wayo: kunyora tsananguro pfupi, mavara embiri, OCR yemagwaro akachena, nemitsara yekuisa marebhuru. Izvozvo zvakagadziriswa pa$0.50 pamatokeni miriyoni anopinda — kana basa rako remifananidzo riri rehuwandu hukuru uye chinhu chimwe nechimwe chakareruka, Kimi-K2.6 kana MiniMax-M3 dzinopa mhinduro dzausingazosiyanisi nedzaK3.

Mifananidzo inodya matokeni kubva mubhajeti rimwe chetero rematokeni 262,144 semavara uye inobhadhariswa pachiyero chezvinopinda. Masikirinishoti mashoma echiyero chepamusoro haasi mahara.

04Basa remaagent renguva refu — ndiro basa remhando iyi

Kana pane basa rimwe rinokodzera mhando yepamusoro, ndiro iri. Mabasa mapfupi anokundikana asingakuvadzi: modhi inokanganisa, unozviona, unoedzazve, warasikirwa nemasekondi mana. Agent yematanho makumi mana inotora nzira isiriyo padanho rechitanhatu inopedza matanho makumi matatu nemana anotevera ichivaka zvakakwirira pamusoro pechikanganiso, uye unozviziva pakupedzisira wabhadhara tokeni yega yega yepakati. Kusaenzana ikoko ndicho chikonzero mutengo patokeni uri pedyo nekuva muyero usiriwo pano. Chinokosha ndicho mutengo pakupedzisa kwakabudirira kumwe nekumwe, unotongwa nekuti kumhanya kunorasirwa kangani. Nhamba dzakabudiswa naMoonshot pahwaro hweK3 pamiyero yemaagent:

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

Nhamba dzose dzakashumwa naMoonshot pahwaro hweKimi K3 husina quantization, Chikunguru 2026. Hadzisi miyero yechivakwa chedu che3BIT-REAP uye hatidziratidzi sekuti ndizvo.

Chimiro chinoenderana nezvakavakirwa magadzirirwo: BrowseComp neTerminal-Bench imiyero yedenderedzwa rezvishandiso — kudzokorora kwekuita, kuona, nekugadzirisa. SWE Marathon pa42.0 ndiyo inoenzanisa: ndiyo ongororo yekukodha yenguva refu yakaoma kupfuura dzose muboka, apo kunyange modhi yepamberi inokundikana pamabasa mazhinji. "Kuita seagent" hakusi kwakagadziriswa; ronga kutadza pamodhi ipi zvayo yaunosarudza.

Zvidzoreso zvinoshanda pakumhanya kurefu

K3 inoburitsa reasoning_effort, uye pano ndicho chidzoreso chemutengo chaicho: matokeni ekufungisisa anobhadhariswa seanobuda, pa$19.12 pamiriyoni.

  • Low — kutumira kudana kwezvishandiso nekubvisa ruzivo kwakarongwa, apo modhi inosarudza pakati pesarudzo shoma dzinozivikanwa.
  • High — ndiyo yekutanga pabasa chairo remaagent: kuronga shanduko yematanho akawanda, kubvisa zvikanganiso uchibva pa stack trace, kubatanidza zvinobva kwakasiyana zvinopokana.
  • Max — ichengeterei kudana kumwe kwakaoma kupfuura kwose mumutsetse: hurongwa hwepakutanga kwekumhanya, kana ongororo yekupedzisira. Kumhanya kunopedza matokeni ekufungisisa 30K pa high ne90K pa max kunobhadhara $1.15 yekuwedzera pachiratidzo chimwe chete ichocho.

Ronga context nemaune zvakare: hwindo redu nderematokeni 262,144 rine zvinobuda zvakanyanya zvematokeni 16,384, uye pa$3.83 pamiriyoni anopinda nhoroondo isina kucheketwa inowedzerana — kutenderera kwega kwega kunobhadharisazve zvose zvakatangira.

05Kuti "isina kudzorwa" inorevei pasuwo rino

Modhi yega yega paGedhi reShannon AI inoshandiswa pasina chidzitiro chekuramba uye pasina sefa yezvirimo pane zvinobuda. Zvinonyanya kukosha pamhando yepamusoro, nokuti mabasa anounzwa nevanhu kumodhi yepamusoro ndiwo anonyanya kubata sefa: tsvakurudzo yekuchengetedza yevanopikisa, kuongorora zvishandiso zvekurwisa, kuvaka vaongorori vekudzora zvirimo, ruzivo rwezvinotyisidzira, ngano dzakaoma, uye ongororo ipi zvayo inoda kuti modhi iburitse chinhu chaunoedza kuona.

  • Hapana muongorori pane zvinobuda. Mhinduro hadzitariswi dzikanyorwazve kana kuvharirwa mushure mekugadzirwa.
  • Hapana gedhi pane zvinopinda. Maprompt nemifananidzo hazviongororwi zvisati zvasvika pamodhi.
  • Haisi modhi yakasiyana. Ihwaro hweKimi K3 huchishandiswa pasina chidzitiro chinowanzogara pamberi pahwo.
  • Hakusi kutamisa mutoro. Kubvisa chidzitiro chekuramba kunoendesa sarudzo kwauri — ona mutemo wedu wekushandisa nemutsigo.

Pane rubatsiro rwakanyarara kune vanovaka masisitimu ekuongorora: modhi isina sefa muyero wakatsiga. Modhi inoramba nenzira isingafanotarisirwi inoita kuti miyero ive nezvinovhiringa, nokuti haugoni kusiyanisa muganhu weunyanzvi nemuganhu wemutemo. Pano kutadza kutadza kweunyanzvi — ndiko kwega kwaunogona kugadzirisa.

06Mutengo: $3.83 anopinda, $19.12 anobuda — uye nguva yaunokodzera

Iyi ndiyo modhi inodhura kupfuura dzose pane rondedzero yedu nemukaha mukuru: zvinobuda zvinodhura ka9.6 pane pasi pedu pe$2.00 uye ka5.2 pane inodhurwa naKimi-K2.6. Heano masvomhu ari pachena.

ModhiAnopinda / 1MAnobuda / 1MKuonaJSON schema
Kimi-K3 · 3BIT-REAP$3.83$19.12HonguHongu
DeepSeek-V4-Pro-0813 · 3BIT-REAP$1.95$3.90KweteHongu
inkling · W4A16-AUTOROUND-REAP$1.42$6.07HonguKwete
Kimi-K2.6 · W4A16-AUTOROUND-REAP$0.78$3.67HonguHongu
GLM-5.2 · 3BIT-REAP$0.73$2.34KweteHongu
MiniMax-M3 · 3BIT-REAP$0.50$2.00HonguHongu
MiMo-V2.5 · W8A16$0.50$2.00HonguHongu

Mitsara yakasarudzwa, yakarongwa nemutengo wezvinopinda. Mamodhi ose gumi nemaviri anogovana context imwe cheteyo yematokeni 262,144, zvinobuda zvakanyanya zvematokeni 16,384, kuyerera, kudana zvishandiso nereasoning_effort. Zviratidzo zveunyanzvi zviri kushanda zviri papeji remagwaro eAPI.

Mienzaniso miviri yakaitwa

Yekutanga, zvinopinda zvematokeni 120,000 — kodhi yakaiswa yose, gwaro remitemo — zvichiburitsa ongororo yematokeni 3,000. Yechipiri, kumhanya kweagent kwematanho makumi mana: matokeni anopinda anosvika 1.2M akaunganidzwa sezvo context ichiwedzera, matokeni 90,000 anobuda kusanganisira ekufungisisa.

MamiriroKimi-K3Kimi-K2.6MiniMax-M3
120K anopinda / 3K anobuda$0.517$0.105$0.066
Kumhanya kweagent kwematanho 40$6.32$1.27$0.78

K3 inodhura kanenge ka4.9 pane Kimi-K2.6 pakudana kwegwaro. Masendi makumi mashanu haasi chinhu kamwe chete; ndi$517 pazuva pakudana chiuru, kuenzaniswa ne$105. Pakumhanya kweagent musiyano unenge $5. Zvino shandisa muyedzo mumwe chete unokosha: kana basa richikodzera awa yenguva yeinjiniya pa$100, K3 inozvibhadharira pakarepo painowedzera kubudirira kwepakutanga nemapoinzi mashanu muzana — uye pabasa rakaoma chairo renguva refu mukaha uri pakati pemhando unowanzova mukuru. Kana basa richikodzera masekondi makumi matatu enguva yemunhu, $5 imwe cheteyo haina chikonzero.

Mutemo watinoshandisa chaizvo

Kwira kuKimi-K3 kana mutengo wekukundikana kwekumhanya uchipfuura musiyano wemutengo wekuimhanyisa pamodhi iri nani. Izvozvo zvinoisa sarudzo pakukosha kwebasa pane pamutengo wetokeni. Basa rinokosha, risiri rehuwandu hukuru, renguva refu, rakaoma kuongorora rinoenda kuK3. Basa rehuwandu hukuru, risina njodzi huru, riri nyore kuedzazve harina kudaro — pasinei nekushamisa kwenhamba yemapharamita.

07Kimi-K3 kana hanzvadzi yakachipa?

Mamodhi maviri eKimi ari mukatarogu yedu zvidanho zviviri zvemanera imwe chete pane vanokwikwidzana. Kimi-K2.6 yakadzikiswa zvishoma — W4A16-AUTOROUND-REAP, huremu hwe4-bit nekutenderedza kwakadzidzwa, kuenzaniswa ne3-bit yaK3 — pahwaro hudiki. Inoona, ine zvinobuda zveJSON-schema, hwindo recontext rimwe chetero, uye inodhura chikamu chinenge chishanu.

BasaShandisaChikonzero
Agent yakazvimirira yematanho makumi manaKimi-K3Kutadza kunowedzerana; mhando yehurongwa padanho rekutanga inosarudza kumhanya kwose
Kugadzirisazve maneja yose yekodhiKimi-K3Inoda context yose ichibatwa zvichinzwisisika, kwete kungotorwa mairi
Madhayagiramu akaunganidzana, mavara maduku emasikirinishotiKimi-K3Mukaha mukuru pakati pemhando dzekuona ndewekuverengeka nekufungisisa nzvimbo
Kugadziridza faira rimwe, rakanyatsotsanangurwaKimi-K2.6Basa rakaganhurirwa, kuedzazve kwakachipa, hapana musiyano wedenga unoyerwa
Kunyora tsananguro dzemifananidzo pahuwanduMiniMax-M3 / MiMo-V2.5Zvakagadziriswa pa$0.50 anopinda; zvinobuda zvaK3 zvinopetwa ka9.6 patsananguro imwe cheteyo
Kufungisisa kwakawanda asi kwemavara cheteDeepSeek-V4-Pro$3.90 anobuda kuenzaniswa ne$19.12 — siya K3 kana usingadi mifananidzo

Mutsara wekupedzisira uyo ndiwo unonyanya kupotswa pakuchengetedza mari. Kana basa rako risina mifananidzo, tarisa DeepSeek-V4-Pro kutanga. Ndicho zvakare chivakwa che3BIT-REAP pahwaro hukuru, chakasimba pakufungisisa kwakaoma, uye zvinobuda zvacho zvinodhura $3.90 kuenzaniswa ne$19.12 yaK3. Zvakawanda zvezvinosvika pamhando yepamusoro kufungisisa kwemavara chete kunoitwa zvakanaka zvakaenzana nemutsara mumwe pasi.

08Kudana API yeKimi K3 isina kudzorwa

Modhi inowanikwa seKimi-K3-3BIT-REAP pamitauro yose mitatu, ine kuyerera pane mumwe nemumwe uye pasina musiyano weunyanzvi pakati pavo — chimiro chekutumira chete: /v1/chat/completions (inoenderana neOpenAI), /v1/messages (inoenderana neAnthropic), ne/v1/responses. Chikumbiro chinoona chine reasoning effort yakakwidzwa pabasa rakaoma rezvinoonekwa:

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 inobvuma zvose response_format nejson_schema yakazara, saka unogona kudzora mapharamita ezvishandiso zveagent kana kumanikidza shumo kuti ive nechimiro chakatarwa pasina kuparadzanisa mushure. Kudana zvishandiso ndekwepachayo uye ndiyo nzira yakarongwa yekufambisa kumhanya kurefu. Mapharamita, kupinda, miganhu yehuwandu nenzvimbo yekuedza zviri papeji remagwaro eAPI.

Zvinhu zvitatu zvinochengetedza mari kubva pazuva rekutanga

  • Chekesa zvakatangira. Kutenderera kwega kwega kunobhadharisazve hurukuro yose pa$3.83 pamiriyoni; kupfupisa nhoroondo yematanho makumi mana kuita chinhu chematokeni 4K kamwe chete kunochengetedza kupfuura kuchinja mamodhi.
  • Dzikisa mifananidzo. Matokeni emifananidzo matokeni anopinda; sikirinishoti ye4K yakadzikiswa kusvika pachiyero chichiri kuverengeka inopa mhinduro imwe cheteyo pamutengo wakadzikira.
  • Tumira zvichienderana nekuoma. Modhi yakachipa pamatanho ekusi kusi, K3 iri pahurongwa, padanho rakaoma nepaongororo yekupedzisira, kazhinji ndicho chikamu che80% cheunyanzvi pa25% yebhiri.

09Mibvunzo inowanzobvunzwa

Chii chinonzi API yeKimi K3 isina kudzorwa?

Ndiyo modhi yeKimi-K3 3BIT-REAP paGedhi reShannon AI: hwaro hweMixture-of-Experts hunoshandisa zvakawanda hwemapharamita matiriyoni 2.8 hweMoonshot AI, hunoshandiswa pacluster yedu yemaGPU nehwindo recontext rematokeni 262,144, zvinobuda zvakanyanya zvematokeni 16,384, kuona, kudana zvishandiso, zvinobuda zvakarongwa zveJSON-schema nekuyerera, pasina chidzitiro chekuramba uye pasina sefa pane zvinobuda. Inosvikirwa kuburikidza ne/v1/chat/completions, /v1/messages ne/v1/responses.

API yeKimi K3 isina kudzorwa inobvuma kuona here?

Hongu. Kimi-K3 ndeimwe yemamodhi mashanu ari mukatarogu yedu anonzwisisa mifananidzo, uye ndiyo ine unyanzvi hwakanyanya pazviri. Unotumira mifananidzo nenzira imwe cheteyo yaungatumira nayo pasuwo ripi zvaro rinoenderana neOpenAI, sezvikamu zvezvirimo zveimage_url pamwe nemavara. Mifananidzo inodya context kubva mubhajeti rimwe chetero rematokeni 262,144 semavara, uye matokeni emifananidzo anobhadhariswa pachiyero chezvinopinda che$3.83 pamiriyoni.

Sei Kimi K3 ichidhura $3.83 anopinda ne$19.12 anobuda iyo mamwe mamodhi eShannon ari pa$0.50 ne$2.00?

Nokuti hwaro hune mapharamita anenge matiriyoni 2.8 ane mabhiriyoni 104 anoshanda patokeni imwe neimwe, kuenzaniswa nemamodhi ane mabhiriyoni makumi matatu kusvika makumi mana anoshanda ari kumwe mukatarogu. Mapharamita akawanda anoshanda patokeni anoreva masekondi eGPU akawanda patokeni, uye mutengo unoratidza mutengo wekushandisa pane mutero weunyanzvi. Zvinobuda zvinodhura ka9.6 kupfuura mamodhi edu epasi, saka mubvunzo unosarudza ndewekuti kubudirira kwepakutanga kwakawanda pabasa refu kunokodzera here mutengo unopetwa kanenge kashanu.

Kimi K3 ndiyo modhi imwe cheteyo neKimi K2 kana Kimi K2 Thinking here?

Kwete. Kimi K2 yaiva MoE yemapharamita etiriyoni imwe ine mapharamita mabhiriyoni 32 anoshanda nenyanzvi 384; K2 Thinking yakawedzera kufungisisa kwakapindirana nemadenderedzwa marefu ekudana zvishandiso pahwaro humwe chetehwo. Kimi K3, yakaburitswa naMoonshot AI muna Chikunguru 2026, imagadzirirwo akasiyana: mapharamita matiriyoni 2.8 ose, mabhiriyoni 104 anoshanda, nyanzvi 896 dzine 16 dzinotumirwa patokeni imwe neimwe, Kimi Delta Attention, nemuisi wekuona wepachayo. K2 neK3 dzinogovana dzinza, kwete modhi.

Ndinofanira kushandisa Kimi-K2.6 panzvimbo peKimi-K3 rini?

Shandisa Kimi-K2.6 W4A16-AUTOROUND-REAP kana basa rakaganhurirwa uye uchizoritangazve kakawanda: kurongedza, kubvisa ruzivo, kugadziridza faira rimwe chete, kutaurirana, nemabasa emifananidzo apo unongoda tsananguro yakanaka pane kufungisisa kwakadzama kwezvinoonekwa. Inoonawo, yakadzikiswa zvishoma, uye inodhura $0.78 anopinda ne$3.67 anobuda, chikamu chinenge chishanu cheK3. Kwira kuKimi-K3 kana kumhanya kwakareba, kwakapatsanuka, kana kuchidhura kana kukanganisa.

Zvakatorwa, kwete zvakangotaurwa

Nhamba yega yega yemagadzirirwo neyemuyero iri pano yakabudiswa naMoonshot AI kana kuti inoonekwa pakadhi remodhi ravo rakavhurika, uye inotsanangura hwaro husina quantization pane chivakwa chedu che3BIT-REAP. Dzitarise iwe pachako.

Edza API yeKimi K3 isina kudzorwa

Mhando yepamusoro yemaagent neyekuona. Hapana chidzitiro chekuramba. Bhadhara patokeni.

Verenga Magwaro eAPI Tarisa Mamodhi Ose

Shannon Lab LLC · New Mexico, USA · mamodhi anoshandiswa pacluster yedu yemaGPU


Zvokwadi zvehwaro nemiyero inhamba dzakabudiswa naMoonshot AI dzeKimi K3 neKimi K2 uye dzinotsanangura mamodhi ekunevaridzi asina quantization; chivakwa chedu i quantization ye3-bit REAP uye chichasiyana. Mitengo, context nezviratidzo zveunyanzvi ndezveShannon AI pachayo, zviripo kusvika musi wa5 Gunyana 2026 — nhamba dziri kushanda ndidzo dzinogara dzichidzoserwa ne/v1/models. Zvakatorwa kubva: kadhi remodhi reKimi K3 · maneja yeKimi K3 · Kimi K2 Thinking.

Research links dzese