Release Date 2026-07-17(Knowledge cutoff Early 2026)1.0M IN · 944K OUTProprietary ModelAPI availableKimi Free and upPerplexity Pro and up

Kimi K3 is Moonshot AI's multimodal reasoning model for sustained development and knowledge work. It expands the earlier family's scale with different attention and residual structures, and official material targets long tasks involving text, images, video, and tools. AIB currently places document work and writing in the top tier and coding in the upper tier, without guaranteeing accuracy or unattended completion on a particular assignment. Full weights are already published in the official repository under a dedicated model license, so earlier plans for a future release are no longer the current availability description. Consider it for large-scale review and iterative engineering, accounting for substantial self-hosting resources and the environment needed for tool execution and verification.

API cost

Market price (median)
$2.92IN$15OUT
Price range19 providers
$0.182~$3IN$9.5~$18OUT
IN
OUT
10/2311/22
SourceOpenRouterLast updated

Real-world usage

No.333 places0.3% share · 869B tokens/wk
2.0T8/31
1.5T9/7
1.5T9/14
1.4T9/21
1.6T9/28
2.2T10/5
2.2T10/12
2.0T10/19
1.7T10/26
1.9T11/2
1.2T11/9
869B11/16
SourceOpenRouterLast updated

Output Speed

Slowper 2,000 tokens
61tok/s
First Output 1.0stotal: 33.8s
SourceOpenRouterLast updated

User Comparisons

Writing Style
Answer length
Longer side
top 21%
Vocabulary diversity
Plainer side
bottom 35%
Sentence rhythm
Steadier side
bottom 26%
Language mixing
Average
top 46%
List usage
Often
21% of lines
Heading usage
Occasionally
13% of lines
Reasoning share
Very heavy
74% of total
Response Speed
instability ±107%
First response 31.4sTotal 66.0s
47 tok/s
16.0ms per token
SourceAIB Inc.Last updated

Expert Performance Evaluation

17.4Reasoning effort: Low148865.079.243.6

More about AA Intelligence Index v5

46.0%28.3%$5,16584.2%36.075.227.6%

By effortLow 41.6%Max 46.0%

More about τ³-Banking

88.7%78.7148947.8%54.3%39.8%53.1%71.6%33.4%46.2%12.9%49.4%72.578.760.0%19.0%12.9%

By effortLow 79.3%Max 88.7%

More about Long Context Reasoning

147150.6%91.2%56.0%79.81.4%95.0%

More about Arena Text Factuality

12.6%153762.2152485.0%76.2165959.5%53.144.2%68.5%82.6%152381.52.0%85.0%25.9%54.4%48.3%31.5%

By effortLow 12.6%Max 12.6%

More about Terminal-Bench 4.0

147515021460148985.5131988.0%146171.4

More about Arena Writing, Literature & Language

46.9%23.4%150293.5%97.2%84.472.2%39.0%87.0%95.0%1517151190.790.70.0%94.5%60.4%60.7%37.0%26.0%91.2%52.7

By effortLow 25.0%Max 46.9%

More about Humanity's Last Exam

34.2%Reasoning effort: Max29.5%91.3%81.6%

More about Furniture Assembly

Specialized fields

14541512151615281512150696.7%90.3%97.0%98.4%85.3%79.965.6%92.0%91.3%96.1%53.0%84.3%90.1%82.0%76.5%

More about Arena Korean

OddBench

74.0%Reasoning effort: Extra High93.6%

By effortMinimal 72.0%Extra High 74.0%

More about BullshitBench v2

See each evaluation’s detail page for its source and last updateData corrections & removal requests

Public Opinion

Users celebrate Kimi K3's multi-agent performance against rivals like Claude, but criticize its steep subscription pricing, rapid token consumption, and frequent server overloads.

Divided223 opinionsYouTube176Bluesky19HF28

Overall rating

IntelligenceVersatilitySpeedAutonomyValue
Kimi K3
IntelligenceReasoning, coding, science, benchmarks72
Praised · 5 of 13 · thin evidence
VersatilityWriting, translation, images and more59
Mostly good · 4 of 7 · thin evidence
ValuePrice and usage limits50
Mixed · 8 of 42
SpeedResponse and task speed36
Mostly poor · 4 of 5 · thin evidence
AutonomyLong tasks and agent runs33
Mostly poor · 5 of 8 · thin evidence

Reaction Summary

1.

Performance & quality

Slightly positive
45 opinions

Users praise Kimi K3's multi-agent swarm and impressive reasoning power, but criticize sluggish inference speeds, frequent server overloads, and the massive hardware requirements needed for local deployment.

2.

Pricing & plans

Slightly negative
42 opinions

Many criticize Kimi K3 for rapid token consumption, high subscription costs reaching hundreds of dollars, and harsh rate limits, noting that the free tier has become practically unusable.

3.

Comparison with rivals

Divided
41 opinions

Opinions are divided against rivals like Claude and DeepSeek: while some consider Kimi K3 a worthy competitor, critics accuse it of being a slower, distilled copy of Anthropic's flagship models.

Actual opinionsselected excerpts

  • استخدمته في فحص كتاب تاريخي وفي البداية قلت له انك باحث وخبير في التاريخ والمطلوب منك التحقق من صحة السرد التاريخي والبحث عن الاخطاء والهفوات في الكتاب والنواقص والرد كان اكثر من رائع بحيث اعطاني 7 ملاحظات لم اتوقعها في الكتاب، بدايتها (تقييمي العام أولاً: الكتاب عمل سردي متين صُنع بصنعة واضحة، وأقوى ما فيه — بصراحة — الملحق المنهجي الذي اعترف بإشكالية سيف بن عمر والواقدي، وهذا أرقى مما في كثير من كتب الفتوح المتداولة. لكن دقّة الملحق تزيد مسؤولية المتن، فوجدت فيه ما يلي: ) وهكذا.. الحقيقة النموذج قوي جدا حيث سبق اني اعطيت نفس الامر للكلاود وما اعطاني مثل هذه الملاحظات
    YouTube@bunoor71View original
  • Great model! For those wondering about hardware: Kimi K3 needs ~1750GB VRAM (22× H100 80GB (MXFP4 native) 或 8× H100 80GB (tensor+expert)). Interactive VRAM calculator: https://openmodelmap.com/kimi-k3 Full deploy guide: https://openmodelmap.com/model/moonshotai/Kimi-K3 Chinese guide: https://openmodelmap.com/kimi-k3/zh
    Hugging Faceduola153View original
  • Kimi K3 é um modelo chinês incrível que desbanca os modelos estadunidenses por um preço muito menor. Produz tudo com qualidade.
    YouTube@aidoluiz3View original
  • Cool to see things like this. Even if its 0.32 tokens a second the fact it runs on that is insane
    Hugging Face8BitStudio3View original
  • انت ادمان، شريت شاشه للسياره من بعد ما بلشت شوف فيديوهاتك، و شريت سيرفر واشتغلت كل يالي عملته بفيدوهاتك السابقه وتصديقا لكلامك عن KIMI جربته الشهر الماضي لحل بعض مشكلات السيرفر يالي عملتو شي خيالي بعد معاناتي مع شات جي بي تي وجيميناي انصح كل متابعينك يستخدموا KIMI انت رائع وفيدوهاتك صادقه وواقعيه
    YouTube@Haloaoooo1View original
  • К3 в бесплатной версии исправил мне ошибки кодов за дипсиком и мои сканеры заработали, я охренел
    YouTube@ДавидКодзасов-п5о0View original
  • What could it be a distillation of? It is better than Opus 4.8, and the Claude 5 models (Opus 5 and Fable 5) return a summary, not the actual CoT, in requests, so they cannot be distilled.
    Hugging Facepython-processing-unit1View original
  • Kimi K3 with the swarm is so nice
    YouTube@paulyflynn3View original
  • Kimi k3 is amazing. I think in the next 3-4 years it will be possible to run the full model at home on a mac studio. That will truly be magical (and expensive). I will definitely be doing that! Sadly it will probably cost 30k for the computer alone. Lol
    YouTube@boot-strapper0View original
  • شكراً لكم على هذه المراجعة القيمة وأتمنا الحصول على باقة kimi الشهرية لأن عندي مشاريع برمجية لا زالت عالقة وتكاليف النماذج الغربية كبيرة.
    YouTube@AhmedSaemAI1View original
SourceYouTubeBlueskyHugging FaceLast analyzed