API na Kimi K3 Marar Tantancewa
Matakin mafi iyawa a Ƙofar Shiga ta Shannon AI: tushe mai hanyoyi da yawa na sigogi tiriliyan 2.8 da gani, kiran kayan aiki da mahalli 262K — babu shimfidar ƙi, kuma farashi mafi tsada a jerinmu. Ga lokacin da wannan farashin ya cancanci a biya shi.
A Taƙaice
Kimi-K3-3BIT-REAP shi ne samfuri mafi ƙarfi da muke hidimarsa kuma mafi tsada: $3.83 kowane tokens miliyan na shigarwa, $19.12 kowane miliyan na fitarwa. Yana da gani, mahalli na tokens 262,144, matsakaicin fitarwa na tokens 16,384, kiran kayan aiki, fitarwar JSON-schema, reasoning_effort da za a iya daidaitawa da streaming a kan dukkan yaruka uku na API — ba tare da shimfidar ƙi ba kuma ba tare da matatar fitarwa ba. Zaɓi ne daidai ga gudun wakilai masu dogon zango, tunani mai wuya mai matakai da yawa, da aikin hoto inda ƙananan bayanan gani suke da muhimmanci. Zaɓi ne mara kyau ga rarrabewa, cirar bayani, hira, da duk abin da kake gudanarwa dubu ɗari a rana.
An rubuta yawancin jagororin zaɓen samfuri kamar dai iyawa kyauta ce. Ba haka ba ne. Tambaya mai muhimmanci game da samfuri na matakin sama ba "shin ya fi kyau" ba ce — yawanci ya fi — sai dai "shin ya fi kyau har ya isa, a wannan aikin, ya cancanci takardar biya sau biyar fiye da samfurin da yake mataki guda ƙasa." Wannan jagorar tana amsa hakan da ainihin lissafi, kuma tana faɗi a fili inda zaɓuɓɓuka masu rahusa suke cin nasara.
01Menene Kimi K3 da gaske
Kimi K3 shi ne babban samfurin buɗaɗɗen nauyi na Moonshot AI, an fitar da shi a Yuli 2026 ƙarƙashin Lasisin Kimi K3 — tsara ta uku ta zuriyar da ta fara da Kimi K2, kuma sabon gine-gine ne da gaske ba wani gyara mai laushi ba.
Tsarin da aka wallafa shimfidu 93 ne — 69 suna amfani da Kimi Delta Attention, 24 suna amfani da gated multi-head latent attention, ɗaya mai kauri — da router na Stable LatentMoE wanda yake zaɓar 16 daga cikin masana 896 a kowane token tare da 2 masu haɗin gwiwa. Moonshot ya bayar da rahoton kusan ingantawa sau 2.5 a ingancin faɗaɗa girma idan aka kwatanta da Kimi K2 daga wannan haɗin. Gani na asali ne ba wani abu da aka manne ba: mai rufa-rufa na MoonViT-V2 mai sigogi 401M yana ciyar da gindin guda wanda yake ɗaukar rubutu.
Lambar tsattsauri tana da kyau a fahimce ta sarai: K3 yana kunna sigogi biliyan 104 a kowane token, idan aka kwatanta da kewayon biliyan 30 zuwa 40 na yawancin katalogin mu. Wannan kusan sau uku ne na lissafin da ake yi ga kowane token da aka samar, kuma shi ne gaba ɗaya dalilin da ya sa wannan samfurin yake kan gaba a jerin farashi. Ba kana biyan kuɗin matakin suna ba ne — kana biyan kuɗin daƙiƙun GPU.
Zuriyar, a taƙaice
- Kimi K2 — sigogi tiriliyan 1, biliyan 32 masu kunnuwa, masana 384 da 8 ake turawa a kowane token, an riga an horar da shi a kan tokens tiriliyan 15.5 ta amfani da mai inganta MuonClip, an gina shi don dakatar da fashewar kulawa da take lalata horon MoE a wannan girman.
- Kimi K2 Thinking — tushe guda da aka koya masa haɗa tunani da amfani da kayan aiki, an bayar da rahoton yana ci gaba da kiran kayan aiki 200 zuwa 300 a jere a gudu guda, da kashi 44.9% a Humanity's Last Exam da kayan aiki da kashi 60.2% a BrowseComp.
- Kimi K3 — an sake gina shi da hanyoyi da yawa tun asali da taga na tokens miliyan a wurin masu shi, tare da horo mai sanin matsawa tun daga matakin gyara mai kulawa zuwa gaba (nauyi na MXFP4, kunnuwa na MXFP8). Shi ya sa K3 yake jure matsawa mai tsauri fiye da samfuran da ba a taɓa horar da su da hakan a rai ba.
Zaren da ya ratsa duka shi ne wakilci mai dogon zango: an inganta kowace tsara don ci gaba da haɗin kai a kan matakai ɗaruruwa maimakon don cin ma'aunan amsa gajera. Wannan ita ce iyawar da kake saya.
02Menene ma'anar "3BIT-REAP", kuma me ya sa samfuri na 2.8T yake buƙatarsa
Ƙarin da ke kan alamar samfurin ba ado ba ne — yana faɗin abin da aka yi wa nauyin kafin a ɗora su a kan gungun GPU namu, kuma shi ne gargaɗin gaskiya da ke manne da kowane iƙirari a nan. Ka fara da lissafi. Samfuri mai sigogi tiriliyan 2.8 a daidaiton bit 16 kusan terabytes 5.6 na nauyi ne. A bit 3 kowane nauyi ya kusanci terabytes 1.05 — har yanzu babba ne, amma tura shi ya yiwu da sauran wuri ga ma'ajiyar KV na tokens 262K. REAP shi ne rabi na biyu: hanyar datsa masana wadda take nazarin waɗanne masana ne router yake kai wa da gaske sannan ta cire waɗanda ba a kasafai ake zaɓarsu ba kafin matsawa maimakon bayanta. A kan samfuri da yake turawa 16 kaɗai daga cikin masana 896 a kowane token wannan ya dace sosai — rarraba mai wannan tsattsaurin yana da dogon jela wanda yake bayar da gudummawa kaɗan ƙwarai ga ainihin zirga-zirga — kuma datsawa da farko yana nufin ana kashe sauran kasafin bit a kan masanan da suke da muhimmanci.
A yi gaskiya game da cinikin
Matsawa tana kashe daidaito, kuma ba daidai wa daida ba. Tana bayyana da farko a tunawa da bayanan dogon jela da kuma a ƙarshen maki na daidaiton lissafi mai wuya, kuma ba sosai a bin umarni, tsara kiran kayan aiki, ko tsarin dogon shirin wakili ba. Horon Kimi K3 mai sanin matsawa shi ne dalilin da ya sa wannan ginin yake tsayawa. Amma idan aikinka shi ne dawo da bayanan da ba a saba da su ba a harbi guda, kada ka ɗauka cewa samfuri na 2.8T da aka matsa ya fi ƙaramin samfuri na 8-bit bisa ƙidayar sigogi kaɗai.
Katalogin mu yana amfani da ƙarin guda uku a matsayin tsani da gangan. 3BIT-REAP yana zuwa a kan manyan tushe, inda babu wani abu da zai shiga. W4A16-AUTOROUND-REAP shi ne nauyi na 4-bit da kunnuwa na 16-bit da aka matsa ta amfani da AutoRound — zagayen da aka koya wanda yake dawo da daidaito na gaske idan aka kwatanta da zagayen zuwa mafi kusa mara nazari — tare da datsawa guda. W8A16 shi ne matakin mafi aminci, ga tushe da ya yi ƙanƙanta har ba ya buƙatar matsawa mai tsauri. A taƙaice: ƙarin yana gaya maka nawa daga cikin na asali ya tsira; farashin yana gaya maka nawa yake kashewa a gudanar da shi.
03Gani: me fahimtar hoto take saya maka
Kimi-K3 yana ɗaya daga cikin samfura biyar masu gani a katalogin mu na samfura goma sha biyu, kuma shi ne mafi iyawa a cikin biyar. Yawancin API na gani na kasuwanci suna gudanar da mai rarrabewa a kan hotunan da aka ɗora sannan su ƙi babban rukunin abubuwan shigarwa kafin samfurin ya taɓa ganin su — abu ne mai ma'ana ga mai amfani na yau da kullum kuma babban cikas ga aiki na halal. Masu binciken tsaro da suke ɗaukar hoton allo na fitarwar amfani da rauni, ƙungiyoyin tantancewa da suke gina masu rarrabewa nasu, aikin bincike na shari'a da duba tarin bayanai, da kimantawar ƙungiyar ja duk suna cin karo da wannan katangar. Kimi-K3 ba shi da shimfidar ƙi kuma ba shi da matata a kan fitarwa.
Inda ganin K3 yake cin farashinsa a kan samfuran gani masu rahusa
- Zane-zane da tsare-tsare masu cunkoso — zanen gine-gine, tsarin hanyoyin sadarwa, da'irorin lantarki: duk inda amsar ta dogara ga bin layuka da alamomi maimakon ambaton abubuwa.
- Ƙaramin rubutu a hotunan allo — alamun tarin kira, fitarwar tasha, ƙwayoyin maƙerin lissafi. Giɓi tsakanin matakan gani ya fi girma a wajen karantuwa a ƙaramin girma.
- Hoto cikin dogon rubutu — hoton allo da aka saka cikin mahalli na tokens 200K, inda dole samfurin ya danganta abin da yake gani da abin da ya karanta tokens dubu hamsin da suka gabata.
- Kwatanta hotuna da yawa — sigogi biyu na wata mahaɗa, ko wani zane idan aka kwatanta da teburin da ya samar da shi.
Inda kuwa ba ya cin farashinsa: rubuta bayanin hoto, rubutun madadin, karanta rubutu daga takardu masu tsafta, da bututun sanya alamomi. An warware waɗannan a $0.50 kowane tokens miliyan na shigarwa — idan aikin hotonka yana da yawa kuma kowane abu mai sauƙi ne shi kaɗai, Kimi-K2.6 ko MiniMax-M3 suna samar da amsoshin da ba za ka bambanta su da na K3 ba.
Hotuna suna cinye tokens daga kasafin tokens 262,144 guda kamar rubutu kuma ana lissafin su a ƙimar shigarwa. Ɗimbin hotunan allo masu ƙuduri sosai ba kyauta ba ne.
04Aikin wakilai mai dogon zango — abin da wannan matakin yake nan dominsa
Idan aiki guda ya ba da hujjar matakin sama, wannan shi ne. Gajerun ayyuka suna faɗuwa da gafara: samfurin ya kuskure, ka lura, ka sake gwadawa, ka rasa daƙiƙa huɗu. Wakili mai zagaye arba'in wanda ya kama reshe mara kyau a zagaye na shida yana kashe sauran zagaye talatin da huɗu yana gina abu mai faɗi a kan kuskure, kuma ka gano hakan a ƙarshe bayan ka biya kowane token da ke tsakani. Wannan rashin daidaito shi ne ya sa farashi a kowane token ya kusa zama ma'auni mara kyau a nan. Abin da yake da muhimmanci shi ne kuɗi a kowace kammalawa mai nasara, wanda yawan lokutan da ake jefar da gudu yake mamayewa. Ga lambobin da Moonshot ya wallafa ga tushen K3 a kan ma'aunan wakilai:
Duk lambobin Moonshot ne ya bayar da rahotonsu ga tushen Kimi K3 marar matsawa, Yuli 2026. Ba ma'aunan ginin mu na 3BIT-REAP ba ne kuma ba ma gabatar da su a haka ba.
Siffar ta dace da abin da aka gina gine-ginen dominsa: BrowseComp da Terminal-Bench ma'aunan zagayen kayan aiki ne — zagayen aikatawa, dubawa, gyarawa da ake maimaitawa. SWE Marathon a 42.0 shi ne mai daidaita ma'auni: kimantawar rubuta lamba mai dogon zango mafi wuya a cikin jerin, inda ko samfurin gaba yake faɗuwa a yawancin ayyuka. "Na wakilai" ba rukuni ne da aka warware ba; ka tsara don gazawa duk samfurin da ka zaɓa.
Saituna masu amfani ga dogayen gudu
K3 yana buɗe reasoning_effort, kuma a nan lever ne na gaske na kuɗi: ana lissafin tokens na tunani a matsayin tokens na fitarwa, a $19.12 kowane miliyan.
- Ƙasa — turawar kiran kayan aiki da cirar bayani mai tsari, inda samfurin yake zaɓa tsakanin ƴan zaɓuɓɓuka da aka sani.
- Sama — na asali ga ainihin aikin wakilai: shirya canji mai matakai da yawa, gyara kurakurai daga alamar tarin kira, da daidaita majiyoyi masu saɓani.
- Matsakaicin ƙarshe — ka ajiye shi ga kira guda mafi wuya a cikin bututu: shirin da ke kan gaban gudu, ko dubawa ta ƙarshe. Gudun da yake kashe tokens 30K na tunani a sama da 90K a matsakaicin ƙarshe yana biyan ƙarin $1.15 saboda wannan alamar guda.
Ka kasafta mahalli da gangan kuma: tagarmu ita ce tokens 262,144 da matsakaicin fitarwa na tokens 16,384, kuma a $3.83 kowane miliyan na shigarwa tarihi da ba a datse shi ba yana taruwa — kowane zagaye yana sake lissafin dukkan farkon.
05Menene "marar tantancewa" yake nufi a wannan ƙarshen hanyar
Ana hidimar kowane samfuri a Ƙofar Shiga ta Shannon AI ba tare da shimfidar ƙi ba kuma ba tare da matatar abun ciki a kan fitarwa ba. Yana da muhimmanci sosai a matakin sama, saboda ayyukan da mutane suke kawo wa samfuri na matakin sama su ne suka fi yiwuwar tarwatsa matata: binciken tsaro na adawa, nazarin kayan aikin kai hari, haɓaka masu rarrabewa na tantancewa, bayanan sirri kan barazana, almara mai wuya, da duk wata kimantawa da take buƙatar samfurin ya samar da abin da kake ƙoƙarin ganowa.
- Babu mai rarrabewa na fitarwa. Ba a duba amsoshi sannan a sake rubuta su ko a toshe su bayan an samar da su ba.
- Babu ƙofar shigarwa. Ba a tantance umarni da hotuna kafin su isa ga samfurin ba.
- Ba wani samfuri dabam ba ne. Wannan shi ne tushen Kimi K3 ana hidimarsa ba tare da shimfidar da yawanci take zama a gabansa ba.
- Ba mika alhaki ba ne. Cire shimfidar ƙi yana mayar da hukuncin gare ka — duba manufar amfani da alhaki tamu.
Akwai wata riba mai natsuwa ga duk wanda yake gina tsarin kimantawa: samfuri marar matatawa tushe ne tsayayye na kwatanci. Samfuri da yake ƙi ba tare da tsari ba yana sa aunawa ta zama mai hayaniya, saboda ba za ka iya raba iyakar iyawa da iyakar manufa ba. A nan gazawa gazawar iyawa ce — nau'in kaɗai da za ka iya gyarawa.
06Farashin: $3.83 shiga, $19.12 fita — da lokacin da ya cancance shi
Wannan shi ne samfuri mafi tsada a jerinmu da tazara mai fadi: fitarwa tana kashe sau 9.6 na ƙasanmu na $2.00 kuma sau 5.2 na abin da Kimi-K2.6 yake kashewa. Ga lissafin a fili.
| Samfuri | Shiga / 1M | Fita / 1M | Gani | JSON schema |
|---|---|---|---|---|
| Kimi-K3 · 3BIT-REAP | $3.83 | $19.12 | Eh | Eh |
| DeepSeek-V4-Pro-0813 · 3BIT-REAP | $1.95 | $3.90 | A'a | Eh |
| inkling · W4A16-AUTOROUND-REAP | $1.42 | $6.07 | Eh | A'a |
| Kimi-K2.6 · W4A16-AUTOROUND-REAP | $0.78 | $3.67 | Eh | Eh |
| GLM-5.2 · 3BIT-REAP | $0.73 | $2.34 | A'a | Eh |
| MiniMax-M3 · 3BIT-REAP | $0.50 | $2.00 | Eh | Eh |
| MiMo-V2.5 · W8A16 | $0.50 | $2.00 | Eh | Eh |
Jeruruwan da aka zaɓa, an tsara su bisa farashin shigarwa. Dukkan samfura goma sha biyu suna raba mahalli guda na tokens 262,144, matsakaicin fitarwa na tokens 16,384, streaming, kiran kayan aiki da reasoning_effort. Alamun iyawa kai tsaye suna kan shafin takardun API.
Misalai biyu da aka yi aiki a kansu
Na farko, shigarwa ta tokens 120,000 — zubar da tarin lamba, ko takardar bayanin doka — tana samar da nazari na tokens 3,000. Na biyu, gudun wakili mai zagaye arba'in: kimanin tokens miliyan 1.2 na shigarwa a tare yayin da mahalli yake taruwa, tokens 90,000 na fitarwa haɗe da tunani.
| Yanayi | Kimi-K3 | Kimi-K2.6 | MiniMax-M3 |
|---|---|---|---|
| 120K shiga / 3K fita | $0.517 | $0.105 | $0.066 |
| Gudun wakili mai zagaye 40 | $6.32 | $1.27 | $0.78 |
K3 yana kashe kusan sau 4.9 na Kimi-K2.6 a kiran takardar. Sisin hamsin ba komai ba ne sau ɗaya; amma $517 ne a rana a kiraye-kiraye dubu, idan aka kwatanta da $105. A gudun wakili bambancin kusan $5 ne. Yanzu ka yi amfani da jarabawa guda da take da muhimmanci: idan aikin ya cancanci awa guda na lokacin injiniya a $100, K3 yana biyan kansa a lokacin da ya ɗaga nasarar wucewa ta farko da maki kashi biyar — kuma a kan aiki mai dogon zango mai wuya da gaske giɓin tsakanin matakan yakan fi hakan girma. Idan aikin ya cancanci daƙiƙa talatin na lokacin kowa, wannan $5 ɗin ba shi da kariya.
Ƙa'idar da muke amfani da ita da gaske
Ka ɗaga zuwa Kimi-K3 sa'ilin da kuɗin faɗuwar gudun ya zarce bambancin farashin gudanar da shi a kan samfuri mafi kyau. Wannan yana sanya shawarar a kan darajar aikin maimakon farashin token. Aiki mai daraja mai girma, mai ƙarancin yawa, mai dogon zango, kuma mai wuyar tabbatarwa yana zuwa K3. Aiki mai yawa, mai ƙarancin haɗari, mai sauƙin sake gwadawa ba ya zuwa — komai girman ƙidayar sigogi.
07Kimi-K3 ko ɗan uwa mai rahusa?
Samfuran Kimi biyu a katalogin mu matakai biyu ne na tsani guda maimakon masu fafatawa. An matsa Kimi-K2.6 da ƙarancin tsauri — W4A16-AUTOROUND-REAP, nauyi na 4-bit da zagayen da aka koya, idan aka kwatanta da 3-bit na K3 — a kan ƙaramin tushe. Yana da gani, fitarwar JSON-schema, taga mahalli guda, kuma yana kashe kusan kashi ɗaya bisa biyar.
| Aiki | Yi amfani da | Me ya sa |
|---|---|---|
| Wakili mai cin gashin kansa mai zagaye arba'in | Kimi-K3 | Gazawa tana taruwa; ingancin shiri a zagaye na ɗaya yake yanke shawarar gudun |
| Sake tsara tarin lamba gaba ɗaya | Kimi-K3 | Yana buƙatar riƙe cikakken mahalli a haɗe, ba kawai a dawo da bayani daga cikinsa ba |
| Zane-zane masu cunkoso, ƙaramin rubutun hoton allo | Kimi-K3 | Giɓin gani mafi girma tsakanin matakan shi ne karantuwa da tunani a kan wuri |
| Gyaran fayil guda, an bayyana shi sosai | Kimi-K2.6 | Aiki mai iyaka, sake gwadawa mai rahusa, babu bambancin rufi da za a auna |
| Rubuta bayanin hoto a yawa | MiniMax-M3 / MiMo-V2.5 | An warware shi a $0.50 shiga; fitarwar K3 tana sau 9.6 don bayani guda |
| Mai nauyi a tunani amma rubutu kaɗai | DeepSeek-V4-Pro | Fitarwa ta $3.90 idan aka kwatanta da $19.12 — ka bar K3 idan ba ka buƙatar hotuna |
Jere na ƙarshe shi ne tanadin da aka fi rasa gani. Idan aikinka bai ƙunshi hotuna ba, ka duba DeepSeek-V4-Pro tukuna. Shi ma ginin 3BIT-REAP ne a kan babban tushe, mai ƙarfi a tunani mai wuya, kuma fitarwarsa tana kashe $3.90 idan aka kwatanta da $19.12 na K3. Yawancin zirga-zirgar da take kai wa matakin sama tunani ne na rubutu kaɗai wanda ake yi masa hidima daidai da kyau jere guda ƙasa.
08Kiran API na Kimi K3 marar tantancewa
An buɗe samfurin a matsayin Kimi-K3-3BIT-REAP a kan dukkan yaruka uku, da streaming a kowanne kuma babu bambancin iyawa tsakaninsu — hanyar aikawa kaɗai ce ta bambanta: /v1/chat/completions (mai jituwa da OpenAI), /v1/messages (mai jituwa da Anthropic), da /v1/responses. Buƙatar gani da ƙoƙarin tunani da aka ɗaga don aiki mai wuya na gani:
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 yana goyon bayan duka response_format da cikakken json_schema, don haka kana iya takura gardamomin kayan aiki na wakili ko ka tilasta rahoto ya shiga siffa daidaitacciya ba tare da nazarin bayan gaskiya ba. Kiran kayan aiki na asali ne kuma ita ce hanyar da aka nufa don tuƙa dogayen gudu. Sigogi, tabbatar da ainihi, iyakokin ƙima da filin gwaji kai tsaye suna kan shafin takardun API.
Abubuwa uku da suke tanadin kuɗi tun rana ta farko
- Ka datse farkon. Kowane zagaye yana sake lissafin dukkan tattaunawar a $3.83 kowane miliyan; taƙaita tarihi mai zagaye 40 zuwa abu mai yanayin tokens 4K sau ɗaya yana tanadi fiye da yawancin canje-canjen samfuri.
- Ka rage girman hotuna. Tokens na hoto tokens na shigarwa ne; hoton allo na 4K da aka rage girmansa zuwa wani abu da har yanzu ake iya karantawa yana ba da amsa guda a kan ƙasa.
- Ka tura bisa wuya. Samfuri mai rahusa a kan zagayen yau da kullum, da K3 a kan shirin, reshe mai wuya da dubawa ta ƙarshe, sau da yawa yana ba da kashi 80% na iyawa a kan kashi 25% na kuɗin.
09Tambayoyin da ake yawan yi
Menene API na Kimi K3 marar tantancewa?
Samfurin Kimi-K3 3BIT-REAP ne a Ƙofar Shiga ta Shannon AI: tushen Mixture-of-Experts mai hanyoyi da yawa na Moonshot AI mai sigogi tiriliyan 2.8, ana hidimarsa a kan gungun GPU namu da mahalli na tokens 262,144, matsakaicin fitarwa na tokens 16,384, gani, kiran kayan aiki, fitarwa mai tsari ta JSON-schema da streaming, ba tare da shimfidar ƙi ba kuma ba tare da matata a kan fitarwa ba. Ana iya kaiwa gare shi ta /v1/chat/completions, /v1/messages da /v1/responses.
Shin API na Kimi K3 marar tantancewa yana goyon bayan gani?
Eh. Kimi-K3 yana ɗaya daga cikin samfura biyar a katalogin mu masu fahimtar hoto, kuma shi ne mafi iyawa a cikinsu. Kana aika hotuna kamar yadda za ka aika su zuwa kowane ƙarshen hanya mai jituwa da OpenAI, a matsayin sassan abun ciki na image_url tare da rubutu. Hotuna suna cinye mahalli daga kasafin tokens 262,144 guda kamar rubutu, kuma ana lissafin tokens na hoto a ƙimar shigarwa ta $3.83 kowane miliyan.
Me ya sa Kimi K3 yake kashe $3.83 shiga da $19.12 fita alhali sauran samfuran Shannon suna $0.50 da $2.00?
Saboda tushen yana da kusan sigogi tiriliyan 2.8 da biliyan 104 masu kunnuwa a kowane token, idan aka kwatanta da samfura masu biliyan 30 zuwa 40 masu kunnuwa a wasu wuraren katalogin. Ƙarin sigogi masu kunnuwa a kowane token yana nufin ƙarin daƙiƙun GPU a kowane token, kuma farashin yana nuna kuɗin hidima maimakon harajin iyawa. Fitarwa tana sau 9.6 na farashin samfuranmu na ƙasa, don haka tambayar da take yanke shawara ita ce ko ƙimar nasara mafi girma a wucewa ta farko a kan dogon aiki ta cancanci kusan sau biyar na kuɗin.
Shin Kimi K3 samfuri ɗaya ne da Kimi K2 ko Kimi K2 Thinking?
A'a. Kimi K2 MoE ne mai sigogi tiriliyan 1 da biliyan 32 masu kunnuwa da masana 384; K2 Thinking ya ƙara tunani mai haɗe da dogayen sarƙoƙin kiran kayan aiki a kan tushe guda. Kimi K3, wanda Moonshot AI ya fitar a Yuli 2026, gine-gine ne dabam: jimlar sigogi tiriliyan 2.8, biliyan 104 masu kunnuwa, masana 896 da 16 da ake turawa a kowane token, Kimi Delta Attention, da mai rufa-rufan gani na asali. K2 da K3 suna raba zuriya, ba samfuri ba.
Yaushe zan yi amfani da Kimi-K2.6 maimakon Kimi-K3?
Yi amfani da Kimi-K2.6 W4A16-AUTOROUND-REAP sa'ilin da aikin yake da iyaka kuma za ka gudanar da shi sau da yawa: rarrabewa, cirar bayani, gyaran fayil guda, hira, da ayyukan hoto inda galibi kake buƙatar bayani mai kyau maimakon tunani mai zurfi a kan gani. Shi ma yana da gani, an matsa shi da ƙarancin tsauri, kuma yana kashe $0.78 shiga da $3.67 fita, wato kusan kashi ɗaya bisa biyar na K3. Ka ɗaga zuwa Kimi-K3 sa'ilin da gudu yake da tsawo, mai reshe-reshe, ko mai tsada idan ya faskara.
Moonshot AI ne ya wallafa kowace lambar gine-gine da ma'auni a nan ko kuma tana bayyane a katin samfurinsu a buɗe, kuma tana kwatanta tushen marar matsawa maimakon ginin mu na 3BIT-REAP. Ka duba su da kanka.
Gwada API na Kimi K3 marar tantancewa
Matakin wakilai da gani mafi iyawa. Babu shimfidar ƙi. Biya kowane token.
Karanta Takardun API Duba Dukkan SamfuraShannon Lab LLC · New Mexico, Amurka · ana hidimar samfura daga gungun GPU namu
Bayanan tushe da ma'aunai lambobin da Moonshot AI ya wallafa ne ga Kimi K3 da Kimi K2 kuma suna kwatanta samfuran asali marasa matsawa; ginin mu matsawar REAP ce ta 3-bit kuma zai bambanta. Farashi, mahalli da alamun iyawa na Shannon AI ne, na yanzu tun 5 ga Satumba, 2026 — ƙimomin kai tsaye koyaushe su ne waɗanda /v1/models yake mayarwa. Majiyoyi: katin samfuri na Kimi K3 · ma'ajiyar Kimi K3 · Kimi K2 Thinking.