ComparEdge · The Dossier · Large Language ModelsFile No. LLM-K3 · Verified 17 July 2026
Moonshot AI · open-weight flagship · July 2026

Kimi Large Language Model Review

Kimi llm dashboard screenshotPlate 01 · Kimi interfaceVisit Kimi

The clearest price-for-intelligence story at the frontier, with an asterisk on trust.

Kimi K3 is the clearest price-for-intelligence story at the frontier right now, and it comes with an asterisk on trust rather than on capability. Moonshot AI shipped a 2.8-trillion-parameter open-weight model that lands fourth of 189 on Artificial Analysis's independent Intelligence Index, ahead of Claude Opus 4.8, at API rates that are a fraction of the Opus tier.

Weights: open, Modified MITIntelligence: frontier-adjacentTrust: unresolvedCertifications: none on file
Editorial rating: 4.5 / 5Prices verified: 17 July 2026Platforms: Web
57
AA Intelligence Index, #4 of 189
$2.31
blended price per 1M tokens
62 tok/s
output speed (AA, provider-dependent)
1M
token context, no length bands
Modified MIT
open weights due ~July 27, 2026
01 / 15Quick answer

Quick answer: should you adopt Kimi K3?#

Kimi K3 is Moonshot AI's July 2026 open-weight flagship, and the honest verdict is that it offers the best price-for-intelligence at the frontier with a real asterisk on trust. Moonshot AI was founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin, three Tsinghua schoolmates, is headquartered in Beijing with a Singapore entity (Moonshot AI PTE. LTD.) operating the global platform, and employs roughly 300 people. It is heavily funded: the latest closed round was about $2B at a $20B valuation in May 2026, led by Meituan's Long-Z Investments, with repeat backers Alibaba and Tencent, and reported talks in June 2026 sought up to a $30B valuation ahead of a possible Hong Kong IPO. On the independent Artificial Analysis Intelligence Index, K3 scores 57, ranked fourth of 189, ahead of Claude Opus 4.8 (56) and GPT-5.5 (55) and just behind Claude Fable 5 (60) and GPT-5.6 Sol (59), with an AA Coding Index of 76.2 and an AA Agentic Index of 50.1; its own vendor benchmark sheet is rosier but agrees on the shape. The weights ship under a Modified MIT license (due around July 27, 2026) whose only real restriction is an attribution requirement above 100M monthly active users or $20M monthly revenue. The main reasons for caution: Moonshot publishes no compliance certifications (no SOC 2 or ISO 27001), it defaults to training on your content unless you sign an enterprise agreement, and it carries an unresolved February 2026 allegation from Anthropic that it used thousands of fraudulent accounts to train its models, framed here as an unadjudicated claim. Adopt K3 for cost-sensitive coding and agentic work where you control the data; look to an Opus-tier provider when you need certifications, a no-training-by-default posture, or the strongest hard-science reasoning.

02 / 15The review verdict

The clearest price-for-intelligence story at the frontier, with an asterisk on trust#

Kimi K3 is the clearest price-for-intelligence story at the frontier right now, and it comes with an asterisk on trust rather than on capability. Moonshot AI shipped a 2.8-trillion-parameter open-weight model that lands fourth of 189 on Artificial Analysis's independent Intelligence Index, ahead of Claude Opus 4.8, at API rates that are a fraction of the Opus tier. If you are optimizing for how much model you get per dollar, it is hard to beat.

What should give a careful buyer pause is not the output quality. It is that Moonshot publishes no compliance certifications, defaults to training on your content, and is carrying an unresolved allegation from Anthropic about how it trained. Adopt it for cost-sensitive engineering where you control the data; think harder before you put regulated data through it.

What it actually costs

The full rate card, the five consumer tiers, and the automatic cache discount live on the pricing page.

03 / 15Who builds it

Moonshot AI, the dark side of the moon#

Kimi is the model family; the company is Moonshot AI, in Chinese 月之暗面, "the dark side of the moon," named after the Pink Floyd album. It was founded in March 2023 by three Tsinghua University schoolmates: Yang Zhilin (CEO, previously a researcher at Meta AI and Google Brain), Zhou Xinyu, and Wu Yuxin. It is counted among China's "six AI tigers," the cohort of well-funded Chinese foundation-model startups.

Moonshot is not a one-model shop. Alongside the Kimi chatbot it ships Kimi Researcher, Kimi Agent, Kimi Code, Kimi Audio, and Kimina Prover. The company first drew attention in October 2023 with a chatbot that handled a 200,000 Chinese-character context, and it has been shipping aggressively since: the flagship line moved from K2 to K2.5 to K2.6 to K2.7-Code to K3 inside a year.

Why the structure matters

Headquarters is Beijing, but the corporate structure is split, and that matters for data questions: the global API runs under the Singapore entity, the China platform under the Beijing one. Headcount is roughly 300 as of 2026 per Wikipedia's company infobox, the best available third-party number rather than a vendor-confirmed count.

Company register
FoundedMarch 2023
FoundersYang Zhilin (CEO), Zhou Xinyu, Wu Yuxin
HeadquartersBeijing, China
Team size~300 employees (2026, third-party estimate)
Global entityMoonshot AI PTE. LTD., Singapore
China entityBeijing Moonshot AI Technology Co., Ltd.
RevenueAnnualized revenue topped $200M in April 2026
04 / 15Funding and backers

The money behind the roadmap#

This is the part worth citing, because the money behind Moonshot tells you how seriously to take its roadmap. Every figure below is verbatim as reported, dated, and sourced; none are converted or adjusted.

Anchor valuation · May 2026$20BLatest closed round: led by Meituan's Long-Z Investments, with Tsinghua Capital, China Mobile, and CPE Yuanfeng. TechCrunch counts $3.9B raised over six months.

Forward signal · June 2026The $30B is a forward signal from reporting on talks in progress, not an achieved valuation, so read it as direction rather than fact.

Seed~$60Mvaluation $300MRaised when the company was about 40 people.
February 2024$1Bvaluation $2.5BLed by Alibaba, whose stake was reported around 36%.
August 2024$300Mvaluation $3.3BTencent and Gaorong Capital.
October 2025~$600Mvaluation $3.8B pre-moneyLed by IDG Capital, with Tencent.
January 2026New roundvaluation ~$4.8BReported by CNBC.
May 2026anchor~$2Bvaluation $20BLatest closed round: led by Meituan's Long-Z Investments, with Tsinghua Capital, China Mobile, and CPE Yuanfeng. TechCrunch counts $3.9B raised over six months.
June 2026$1-2B in talksvaluation up to $30B soughtReported talks, not closed; a Hong Kong IPO is being prepared by unwinding the offshore structure.

The anchor number is $20B, closed in May 2026. The $30B is a forward signal from reporting on talks in progress, not an achieved valuation, so read it as direction rather than fact. Notable backers named by TechCrunch include Alibaba, Tencent, HongShan (formerly Sequoia China), ZhenFund, IDG Capital, and 5Y Capital. Two of China's largest tech companies are repeat backers, which is a meaningful signal about staying power even before you get to the benchmarks.

05 / 15Release history

How fast this lineup moves#

A compressed history, because it explains both the shipping pace and which models are already on the way out.

Oct 2023First Kimi chatbot, 200K Chinese-character context
Feb 2024Alibaba-led $1B round
Jan 2026Kimi K2.5 adds native vision (400M-param MoonViT encoder)
Mar 2026Hong Kong IPO reported under consideration
Jun 11, 2026K2.7-Code weights appear on Hugging Face
Jul 15-16, 2026Kimi K3 launches (OpenRouter listing July 16)
~Jul 27, 2026Full K3 open weights due

On the retirement side: the kimi-k2 series was deprecated May 25, 2026, kimi-latest on January 28, 2026, and kimi-thinking-preview on November 11, 2025. The Moonshot V1 models and kimi-k2.5 sunset for new users around August 31, expected 2026.

06 / 15Insider context

The open-weight bet#

The strategy worth understanding is the open-weight release cadence. Moonshot ships flagship weights to Hugging Face weeks after API launch under a Modified MIT license, which turns every self-hosting team into distribution and puts price pressure on closed rivals. K2.6 became the second most-used LLM on OpenRouter in May 2026, so the family had real adoption before K3 landed.

The revenue signal points the same way: annualized revenue topped $200M by April 2026 on TechCrunch's reporting, most of it API. An IPO in Hong Kong is being prepared. A company this funded, growing this fast, with Alibaba and Tencent as repeat backers, is not a research lab experiment; it is a priced-to-win commercial operation.

07 / 15Trust and governance

Where your data lives, and what is missing#

Enterprises and AI agents ask the same three questions here: where does my data live, can I self-host, and is any of this certified. The honest answers are mixed.

Info · data residency
Data residency

The global platform (platform.kimi.ai) is operated by the Singapore entity; its privacy policy states data is stored and processed on servers in Singapore, under Singapore law, with disputes via the Singapore International Arbitration Centre. Explicitly not mainland China. The China platform (platform.kimi.com) runs under the Beijing entity with RMB billing and a 6% China VAT invoice.

Watch · training on your content
Training on your content

By default your content may be used to develop and improve the models. Opting out is only available through enterprise arrangements or separate written agreements. Moonshot does not claim ownership of your content and offers GDPR-style rights, but the default is train-on-your-data, not the reverse.

Strength · license and self-hosting
License and self-hosting

Weights ship under a Modified MIT License (confirmed verbatim on K2 through K2.7, the strong prior for K3): standard MIT permissions, no copyleft, no field-of-use limit. The single modification: products above 100M monthly active users or $20M monthly revenue must display the model name prominently. Below that scale it is effectively permissive commercial use and self-host.

Watch · certifications
Certifications

This is the gap. No SOC 2, ISO 27001, HIPAA, or any other certification is stated anywhere in the terms, privacy policy, or docs. Do not assume one exists. If procurement needs the paperwork, K3 does not currently have it.

08 / 15Independent versus vendor

Two benchmark stories, one shape#

Independent: Artificial Analysis

AA's Intelligence Index (v4.1, nine evaluations) puts K3 at 57, ranked #4 of 189, against a class average of 30. On AA's peer board that places it just behind Claude Fable 5 (60) and GPT-5.6 Sol (59), and ahead of Grok 4.5 (54), GLM-5.2 (51), Claude Opus 4.8 (56) and GPT-5.5 xhigh (55). AA breaks that into a Coding Index of 76.2 and an Agentic Index of 50.1, both top-few among frontier families. There is no standalone AA math index; the reasoning signal shows up in component evals.

Vendor: Moonshot's own sheet

Moonshot's published 40-row benchmark sheet is rosier and should be read as such: Terminal Bench 2.1 at 88.3, BrowseComp at 91.2, GPQA-Diamond at 93.5, DeepSearchQA at 95.0. Its own framing is honest about the ceiling, saying K3 "trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol," while leading the rest of its test set.

When the independent index and the vendor sheet agree on the shape (strong, top-5, behind the two best US frontier models), you can trust the shape. The vendor's specific numbers are the optimistic read; AA is the check.

The working notes

Per-eval sub-scores, the caching mechanics in code, and every API call walked through line by line live on the technical page.

09 / 15The competitive field

Who Kimi K3 actually competes with#

The spine of every K3 decision is the price-for-intelligence ratio, so here is the whole frontier board on one axis: Artificial Analysis's independent Intelligence Index, with the honest money argument next to each name. K3's blended price is about $2.31 per 1M tokens; the two models above it cost several times that.

ModelAA IndexThe money argument
Claude Fable 560The capability ceiling. If the task needs the strongest model available, price stops being the axis.
GPT-5.6 Sol59Second of the two closed leaders K3 explicitly trails; same calculus as Fable 5.
Kimi K3you are here57$3.00 input / $15.00 output per 1M, $0.30 cached input, 1M context, open weights. The price-performance pick of the top five.
Claude Opus 4.856$5.00 / $25.00 per 1M: one point below K3 at a higher rate, but with SOC 2-class paperwork and a no-training default K3 cannot offer.
GPT-5.5 xhigh55Trails K3 on the index; the argument for it is ecosystem, not price.
Grok 4.554Close on intelligence, weaker on the coding and agentic sub-scores where K3 leads.
GLM-5.251The other Chinese open-weight contender; cheaper tier, six points back.
Gemini 3.5 Flash50A speed-and-cost play, not a frontier rival.
DeepSeek V4 Pro44$0.435 / $0.87 per 1M: the price floor. Thirteen points below K3; the trade is intelligence for cost.

Read the board top-down and the placement is plain. Above K3 sit two closed models you pay a premium for; directly below sits Opus 4.8, which loses the index by a point but wins every compliance conversation; far below sits DeepSeek, which wins on price alone. K3 owns the middle: frontier-adjacent scores at open-weight economics, for buyers whose blocker is budget rather than an auditor.

10 / 15Use-case fit

Where the sub-scores say it earns its keep#

Earns its keep · autonomous coding
Autonomous coding

The 76.2 AA Coding Index and a top-6 Terminal-Bench place K3 among the best available for long-running software work: large-codebase analysis, tool coordination, multi-step tasks. The 1M context holds a whole repository in view.

Earns its keep · tool-heavy agents
Tool-heavy agents

The 50.1 AA Agentic Index and a first-place independent finish on tau-cubed Banking tool use make it a credible agent base. The automatic cache keeps a repeated system prompt and tool schema cheap across many turns.

Earns its keep · long-context and vision
Long-context and vision

K3 ranks first on AA's long-context reasoning eval, and native image and video input suit a read-screen, edit-code, check-output loop for UI, games, and CAD work.

Weak spot · hard science reasoning
Hard science reasoning

Tenth of its set on the CritPt physics eval and mid-pack on knowledge accuracy and visual reasoning. This is not the model for frontier physics or pixel-precise multimodal work.

11 / 15Strengths and watch-outs

The honest two-column read#

What it does well
  • Top-5 independent intelligence (AA Index 57) at a fraction of Opus-tier cost
  • Automatic 90% cache-hit discount with nothing to configure
  • 1M context as standard, no context-length bands
  • Open weights (Modified MIT), so self-hosting is a real option
  • Singapore data residency on the global platform
What to watch
  • Reasoning is always on and the model is very verbose, so output cost runs high
  • No Batch discount for K3 at launch
  • No published compliance certifications at all
  • Default is train-on-your-content unless you sign an enterprise deal
  • An unresolved Anthropic allegation hangs over the company (see Incidents)
Reach for it if

Reach for it if you are a cost-sensitive engineering team doing high-volume coding or agentic work, you control the data you send, and you want frontier-adjacent quality without frontier pricing. The open weights make it doubly attractive if you can self-host and want to remove the vendor from the loop entirely.

Skip it if

Skip it if you are in a regulated industry that needs SOC 2 or ISO 27001 on file, if a train-on-your-data default is a non-starter, or if your workload is hard science reasoning where K3 is measurably weaker. In those cases an Opus-tier provider costs more per token but answers the questions your auditor will ask.

The true cost of ownership

The verbosity multiplier, the hidden fees, and the break-even against an Opus-tier model are worked out on the cost guide.

12 / 15Incidents and controversy

Reported neutrally, with the date and the source#

We report this with the date and the source, because balanced sourcing is the point of a trust section.

Exhibit A · Kimi
Filed February 2026
Anthropic accuses Moonshot of training on Claude via fraudulent accounts
Unresolved allegation

Per Wikipedia's entry on the company, Anthropic accused Moonshot of violating its terms of service by using thousands of fraudulent accounts to obtain access and train its own large language models. This is an allegation, not an adjudicated finding. As of this writing it is unresolved: there is no public ruling, settlement, or confirmed outcome, and Moonshot's public response is not documented in the sources we hold. We are not stating it as fact. We are noting that a major US lab publicly accused Moonshot of training against its terms, that the matter is open, and that a buyer weighing trust should factor in an unproven but serious claim from a credible source.

Source · Wikipedia: Moonshot AI

Any further incidents will be added here as they are confirmed.

13 / 15Reception

What third parties said at launch#

SiliconANGLE, July 16, 2026

Framed K3 as "the world's largest open-weights model" and reported benchmarks showing it outperforming the best OpenAI and Anthropic models in some applications, while still trailing GPT-5.6 Sol and Claude Fable 5 in others, described as extremely close.

Arena.ai front-end board, July 16, 2026

A second independent leaderboard ranked K3 above both GPT-5.6 Sol and Claude Fable 5, and 17 places above Moonshot's own K2.6. Arena CEO Anastasios Angelopoulos said on X it "may be the single biggest release of the year" and called it the moment open-source Chinese models surpassed US models. That is an attributable opinion from a named source, not a benchmark result.

Distribution: OpenRouter and Cloudflare

K3 launched on OpenRouter (single provider, Moonshot direct) and has a live model page on Cloudflare Workers AI as of July 17, 2026. Moonshot's prior K2.6 was the second most-used LLM on OpenRouter in May 2026, so the family already has real adoption behind it.

14 / 15Questions on file

Kimi FAQ#

Q1What does Kimi K3 cost?
The API is pay-as-you-go: $3.00 per 1M input tokens on a cache miss, $0.30 on a cache hit, and $15.00 per 1M output, all at a 1M context. There is no free API tier, only a $1 minimum recharge. The separate consumer Kimi app runs from a free Adagio tier up to $199 a month.
Q2Is Kimi K3 open source?
It is open weight rather than open training data. Moonshot said it will publish the full K3 weights around July 27, 2026 under a Modified MIT license, following the same license used on the K2 family. That lets you fine-tune and self-host, though serving a 2.8T model needs serious hardware.
Q3How good is Kimi K3 compared to Claude and GPT?
On Artificial Analysis, the independent benchmark, K3 posts an Intelligence Index of 57, fourth of 189. That is ahead of Claude Opus 4.8 (56) and GPT-5.5 (55), and just behind Claude Fable 5 (60) and GPT-5.6 Sol (59). It leads on long-context and tool-use evals and is weaker on physics reasoning.
Q4Where is my data stored?
It depends which platform you use. The global platform (platform.kimi.ai) is operated by Moonshot AI PTE. LTD. in Singapore, with data stored on Singapore servers under Singapore law. The China platform (platform.kimi.com) runs under the Beijing entity. By default your content may be used to improve the models unless you have an enterprise agreement.
Q5Which Kimi models are available besides K3?
The current catalog runs five deep. kimi-k3 is the 2.8T flagship at $3.00 per 1M input tokens ($0.30 on a cache hit) and $15.00 per 1M output with the only 1M-token context. kimi-k2.7-code handles dedicated coding at $0.95/$4.00 per 1M with the most stable structured output, and its HighSpeed variant doubles the price for roughly 180 tokens per second. kimi-k2.6 is the general multimodal cost-saver at $0.95/$4.00, kimi-k2.5 at $0.60/$3.00 sunsets for new users around August 31, and the legacy moonshot-v1 family is priced per context size on its way out.
Q6What do the Kimi consumer plans include, and which one matters?
Five tiers with musical names: Adagio is free with one agent task and 500MB of storage, Moderato at $19 a month adds paid-tier throughput, Allegretto at $39 doubles agent capacity, Allegro at $99 is the one that matters because the K3 Extra Long Chat Capacity of up to 1M tokens switches on there, and Vivace at $199 maxes the credit allowances. Annual billing trims each paid tier, and the vendor has flagged the lineup as changing soon, with Kimi and Kimi Code benefits due to be separated.
15 / 15Review sources

The record, cited#

  1. Wikipedia: Moonshot AI (company, founders, headcount, February 2026 allegation)en.wikipedia.org/wiki/Moonshot_AI
  2. TechCrunch: $2B raise at $20B valuation, May 2026techcrunch.com/2026/05/07/chinas-moonshot-ai-raises-2b-at-20b-valuation-as-demand-for-open-source-ai-skyrockets
  3. CNBC: ~$4.8B valuation round, January 2026cnbc.com/2026/01/19/alibaba-backed-startup-moonshot-ai-alibabi-backed-startup-chinese-ai-ipo.html
  4. The Decoder: $30B valuation talks, June 2026the-decoder.com/moonshot-ai-targets-a-30-billion-valuation-more-than-six-times-its-late-2025-worth
  5. Moonshot model use terms (residency, training default)platform.kimi.ai/docs/agreement/modeluse.md
  6. Modified MIT license text (K2.6 weights, family prior)huggingface.co/moonshotai/Kimi-K2.6/raw/main/LICENSE
  7. Artificial Analysis: Kimi K3 independent benchmarksartificialanalysis.ai/models/kimi-k3
  8. Moonshot K3 announcement and vendor benchmark sheetkimi.com/blog/kimi-k3
  9. SiliconANGLE launch coverage, July 16, 2026siliconangle.com/2026/07/16/chinas-moonshot-throws-gauntlet-kimi-k3-worlds-largest-open-weights-model
Data syncLast verified: July 17, 2026Revision: ce-rev-20260717-kimi-k3-001✓ Pricing updatedUp to date
ComparEdge · The Dossier · KimiFile LLM-K3 · Verified 17 July 2026Reviewed independently · sources cited above