By Matt Farmer / Published Jul 17, 2026 / Last verified Jul 17, 2026

Kimi K3 explained: China's new AI heavyweight.

Moonshot AI released a Chinese AI model that scores close to GPT-5.6 Sol and Claude Fable 5, costs less per token, and is promised to become open weight on July 27. Here is what that means—without the jargon.

Try Kimi K3 Official API

Matt Farmer beside a futuristic Kimi K3 robot designing a website interface — Matt Farmer's visual breakdown of Kimi K3

Independent score: 57 vs 59 / 60
Context window: 1 million tokens
API price: $3 in / $15 out
Open-weight promise: July 27, 2026

01 / Why it matters

A Chinese AI lab just moved closer to the frontier

The simple version: K3 is not clearly the best model in the world. It is close enough to change the conversation.

Moonshot AI launched Kimi K3 on July 16 with consumer products, coding tools, a paid API, native image understanding, and a one-million-token context window. In ordinary language, it can work with very large amounts of text, see images, and stay involved in longer jobs.

The independent comparison is the important part. Artificial Analysis scored K3 at 57, versus 59 for GPT-5.6 Sol and 60 for Claude Fable 5. A two- or three-point gap does not make the models interchangeable, but it puts K3 in the same serious buying conversation as the current leaders.

The second reason people care is the promised open-weight release. If Moonshot ships the full checkpoint and a workable license on July 27, researchers and companies could inspect and host the model themselves. That would not make a 2.8-trillion-parameter model easy to run, but it would give the wider AI ecosystem far more control than a closed API alone.

02 / The scoreboard

How close is K3 to GPT-5.6 and Claude?

There is no single test that proves which model is best. These two snapshots answer narrower questions, so they stay separate instead of being blended into a made-up winner score.

Independent intelligence index

Artificial Analysis Intelligence Index v4.1, captured July 17, 2026. Higher is better. This is a broad independent composite, not a percentage score.

Artificial Analysis

Claude Fable 560

GPT-5.6 Sol59

Kimi K357

K3 lands within three points of Fable and two points of Sol—a meaningful gap, but a genuinely competitive result.

Show the index values as a table

Independent intelligence index values
Model	Value
Claude Fable 5	60
GPT-5.6 Sol	59
Kimi K3	57

Preliminary Arena WebDev score

Arena WebDev Overall snapshot captured July 16, 2026. K3's 1679 ±17 result was marked Preliminary and can change as more votes arrive.

Arena leaderboard

Kimi K31679 ±17

Claude Fable 51631

GPT-5.6 Sol1618

K3 led this specific web-development snapshot. That is strong evidence for visual coding—not proof that it wins every kind of work.

Show the Preliminary WebDev values as a table

Preliminary Arena WebDev score values
Model	Value
Kimi K3	1679 ±17
Claude Fable 5	1631
GPT-5.6 Sol	1618

Compare the full model field

03 / The price

K3 is cheaper—but output still matters

K3's published API rates are lower than the two frontier models above. The useful question is not just what one token costs; it is how many tokens, turns, retries, and tools the finished job needs.

Published API price per million tokens

The lighter bar is cache-miss input. The darker bar is generated output. Provider caching rules and completed-job token use can change the final bill.

Official Kimi pricing

Kimi K3

Input$3

Output$15

GPT-5.6 Sol

Input$5

Output$30

Claude Fable 5

Input$10

Output$50

Kimi also publishes a $0.30 cached-input rate per million tokens. That rate is not mixed into the bars because provider caching rules are not directly equivalent.

K3 has the lowest published input and output rates in this three-model comparison, but a verbose or retry-heavy job can still cost more than expected.

Show pricing as a table

Published API price comparison per million tokens
Model	Cache-miss input	Output	Cached input
Kimi K3	$3	$15	$0.3
GPT-5.6 Sol	$5	$30	Provider-specific
Claude Fable 5	$10	$50	Provider-specific

Illustrative worksheet

$0.02

Quick chat

One turn; no tools or retries.

Illustrative worksheet

$0.36

Coding with a familiar project

90% of 100K input cached; 20K output/reasoning.

Illustrative worksheet

$1.69

Long research task

75% of 500K input cached; tools and web fees excluded.

Run your own numbersUse all six supplied workload presets or change the token mix, turns, retries, and tool fees.

((cached / 1M × $0.30) + (uncached / 1M × $3.00) + (output / 1M × $15.00)) × turns × retry + tool fees

Preset assumptions: One turn; no tools or retries. Source worksheet total: $0.0210.

Cached input tokens per turnUncached input tokens per turnOutput / reasoning tokens per turnTurnsRetry multiplier

1.0×

Tool fees (USD)

Cached input: $0.0000
Uncached input: $0.0060
Output / reasoning: $0.0150
Base cost per turn: $0.0210

Estimated completed job

$0.0210

Estimate only. Actual caching, generated reasoning, retries, rate limits, taxes, and tool charges can change the bill.

04 / Everyday use

What can ordinary people do with K3?

The launch material is most useful when translated into jobs people recognize. These are the four clearest reasons to try the model, with the evidence kept in proportion.

Coding and visual development

Build websites and software

K3 is designed for long coding sessions, visual website work, debugging, and multi-step agent tasks. Moonshot reports strong results across several coding tests, while the independent composite still places Fable and Sol slightly ahead overall.

1M-token context

Research very long documents

The one-million-token context window gives K3 room for large collections of reports, transcripts, code, and reference material. That does not guarantee perfect recall, but it makes larger research jobs possible without splitting everything into tiny pieces first.

Documents, slides, and spreadsheets

Create office work

Moonshot demonstrates K3 producing documents, presentations, spreadsheets, dashboards, and other editable business material. Treat those examples as a reason to test the workflow—not proof that every spreadsheet formula or presentation will be correct.

Native visual input

Understand images

K3 can accept images as part of a prompt, which matters for visual coding, document analysis, charts, screenshots, and design feedback. Native vision means the model can reason about what it sees instead of relying only on extracted text.

05 / The promise

What “open weight” actually means

This is the most important phrase to get right. Moonshot promised the model files; it had not yet supplied the checkpoint, license, or technical report when this page was verified.

Think of open weights as receiving the engine of the model. A company can potentially inspect that engine, run it on its own hardware, and build around it instead of sending every request to someone else's hosted service.

Open source is a broader promise. The license determines what people are allowed to modify, redistribute, or use commercially, and a technical report explains more of how the model was built. Until those files can be inspected, “open weight” remains the accurate description of Moonshot's promise.

For most people, nothing changes overnight: the easiest path will still be the Kimi website or API. The difference matters most to researchers, governments, and companies that need more control over where a model runs, how it is inspected, and what happens to their data.

Open weight

The trained model files become available under whatever terms the license permits.

Open source

A wider claim involving code, documentation, rights, and the exact license—not just downloadable weights.

06 / The catch

Read this before you choose K3

K3 is impressive because the honest version is already strong. It does not need a universal-winner headline or an open-source label that the evidence cannot yet support.

01
Close is not the same as first
Artificial Analysis scored K3 at 57, behind Fable at 60 and Sol at 59 in the July 17 snapshot. That is a small frontier gap, but it is still a gap.
02
Cheap tokens can become an expensive job
Artificial Analysis measured 130 million output tokens across its index run. Long reasoning, retries, tool calls, and cleanup can eat into K3's lower published rates.
03
Hosted API privacy needs review
This review did not verify zero retention or no-training-by-default for the standard API. Sensitive work needs current terms and, where appropriate, a written enterprise agreement.
04
The open-weight story is still a promise
At the July 17 cutoff, the checkpoint, license, technical report, third-party hosting, and independent self-host results were not yet available to inspect.

Matt's read: K3 belongs on a serious evaluation shortlist today. Use it where long context, visual work, and lower API rates fit the job—and recheck the evidence when the July 27 artifacts arrive.

Try K3 now

You have a real coding, research, document, or visual task and can compare the finished result with your current model.

Compare before switching

Your work is difficult enough that a two- or three-point independent gap could matter more than the lower token rate.

Wait for July 27

Your plan depends on downloading the weights, inspecting the license, or hosting the full model on your own infrastructure.

07 / The receipts

The technical evidence, if you want it

Nothing useful has been thrown away. The complete benchmark matrix, model comparison, architecture, test boundary, community questions, methodology, changelog, and sources live below without blocking the main explanation.

Open the complete technical evidence37 benchmarks, 12 comparison models, six cost presets, architecture, testing notes, privacy, changelog, and full sources.

Full 37-row benchmark explorerFilter by workload, evidence scope, result, and harness while retaining every configuration caveat.

Category

Test scope

K3 result

Harness

Showing 37 of 37 entries

Kimi K3 benchmark explorer with six model columns, harness details, evidence scope, and configuration caveats
Benchmark	K3	Fable 5	Sol	Opus 4.8	GPT-5.5	GLM-5.2	Result	Harness + caveat	Source
Coding / Vendor test / scoreDeepSWETrails	K367.5	Fable 570	Sol73	59	67	46.2	Trails	KimiCode Official leaderboard figures use multiple harnesses; K3 is 67.3 under mini-SWE-agent. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 59 / GPT-5.5 67 / GLM-5.2 46.2	Open evidence
Coding / Vendor test / pass rateProgram BenchLeads	K377.8	Fable 576.8	Sol77.6	71.9	70.8	63.7	Leads	KimiCode GLM score comes from vendor blog; others from Vals AI per Moonshot. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 71.9 / GPT-5.5 70.8 / GLM-5.2 63.7	Open evidence
Coding / Vendor test / scoreTerminal Bench 2.1Trails	K388.3	Fable 584.6	Sol88.8	84.6	83.4	82.7	Trails	KimiCode Other models use best reported harness; not a pure model comparison. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 84.6 / GPT-5.5 83.4 / GLM-5.2 82.7	Open evidence
Coding / Vendor test / dominanceFrontierSWETrails	K381.2	Fable 586.6	Sol71.3	66.7	64.9	67.3	Trails	KimiCode Harness differs; dominance recomputed from raw scores on July 16. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 66.7 / GPT-5.5 64.9 / GLM-5.2 67.3	Open evidence
Coding / Vendor test / scoreSWE MarathonLeads	K342	Fable 535	Sol39	40	14	13	Leads	Claude Code K3/Fable/Opus use Claude Code; Sol uses Codex; GLM from vendor. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 40 / GPT-5.5 14 / GLM-5.2 13	Open evidence
Coding / Vendor test / scorePostTrain BenchTrails	K336.6	Fable 541.4	Sol34.6	34.1	28.4	34.3	Trails	Claude Code K3/Fable/Sol averaged three official Harbor runs; Fable can fall back to Opus. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 34.1 / GPT-5.5 28.4 / GLM-5.2 34.3	Open evidence
Coding / Vendor test / scoreMLS Bench LiteTrails	K348.3	Fable 549.9	Sol46.2	42.8	35.5	40.4	Trails	KimiCode Claude models use Claude Code and GPT models use Codex. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 42.8 / GPT-5.5 35.5 / GLM-5.2 40.4	Open evidence
Coding / Internal suite / internal scoreKimi Code Bench 2.0Trails	K372.9	Fable 576.9	Sol64.8	71.7	69	64.2	Trails	KimiCode and Claude Code In-house Kimi benchmark; all max except GPT-5.5 xhigh. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 71.7 / GPT-5.5 69 / GLM-5.2 64.2	Open evidence
Agentic / Vendor test / EloGDPval-AA v2Trails	K31,668	Fable 51,760	Sol1,748	1,600	1,494	1,514	Trails	Not specified Artificial Analysis source; snapshot can change. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 1,600 / GPT-5.5 1,494 / GLM-5.2 1,514	Open evidence
Agentic / Vendor test / scoreBrowseCompLeads	K391.2	Fable 588	Sol90.4	84.3	84.4	Not reported	Leads	Not specified Context compaction at 300K; K3 scores 90.4 with raw 1M/no context management. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 84.3 / GPT-5.5 84.4 / GLM-5.2 Not reported	Open evidence
Agentic / Vendor test / F1DeepSearchQALeads	K395	Fable 594.2	SolNot reported	93.1	Not reported	Not reported	Leads	Not specified Sparse comparator coverage. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 93.1 / GPT-5.5 Not reported / GLM-5.2 Not reported	Open evidence
Agentic / Vendor test / scoreToolathlon-VerifiedTrails	K373.2	Fable 577.9	Sol74.9	76.2	73.5	59.9	Trails	Not specified Harness details not expanded in launch footnotes. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 76.2 / GPT-5.5 73.5 / GLM-5.2 59.9	Open evidence
Agentic / Vendor test / scoreMCP AtlasTrails	K384.2	Fable 584.7	Sol83.6	83.6	82.8	82.6	Trails	Not specified 500-task public subset; 100-turn limit; Gemini 3.1 Pro judge. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 83.6 / GPT-5.5 82.8 / GLM-5.2 82.6	Open evidence
Agentic / Vendor test / scoreAutomation BenchLeads	K330.8	Fable 529.1	Sol29.7	27.2	22.7	12.9	Leads	Official GitHub setup 600-task public subset. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 27.2 / GPT-5.5 22.7 / GLM-5.2 12.9	Open evidence
Agentic / Vendor test / scoreJob BenchTrails	K352.9	Fable 557.4	Sol46.5	48.4	38.3	43.4	Trails	Not specified Vendor table; audit exact benchmark protocol before strong claims. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 48.4 / GPT-5.5 38.3 / GLM-5.2 43.4	Open evidence
Agentic / Vendor test / EloAA-BriefcaseTrails	K31,548	Fable 51,583	Sol1,495	1,354	1,158	1,260	Trails	Not specified Artificial Analysis source; snapshot can change. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 1,354 / GPT-5.5 1,158 / GLM-5.2 1,260	Open evidence
Agentic / Vendor test / scoreAPEX-AgentsTrails	K337.6	Fable 543.3	Sol39.9	39.4	38.5	35.6	Trails	Not specified Vendor table. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 39.4 / GPT-5.5 38.5 / GLM-5.2 35.6	Open evidence
Agentic / Vendor test / scoreOffice QA ProTrails	K363.3	Fable 569.9	Sol63.2	63.9	60.9	41.4	Trails	Claude Code PDFs rendered as images; GPT models use Codex; starred non-K3 scores in source. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 63.9 / GPT-5.5 60.9 / GLM-5.2 41.4	Open evidence
Agentic / Vendor test / scoreSpreadsheetBench 2Leads	K334.8	Fable 534.7	Sol32.4	31.6	29.1	28.1	Leads	Claude Code GPT models use Codex; starred non-K3 scores in source. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 31.6 / GPT-5.5 29.1 / GLM-5.2 28.1	Open evidence
Agentic / Internal suite / internal scoreDECK-BenchTrails	K373.5	Fable 573	Sol74.7	66.9	68.2	68.6	Trails	Not specified In-house benchmark. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 66.9 / GPT-5.5 68.2 / GLM-5.2 68.6	Open evidence
Reasoning / Vendor test / scoreGPQA-DiamondTrails	K393.5	Fable 592.6	Sol94.1	91	93.5	91.2	Trails	Not specified Vendor table. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 91 / GPT-5.5 93.5 / GLM-5.2 91.2	Open evidence
Reasoning / Vendor test / scoreHLE-FullTrails	K343.5	Fable 553.3	Sol44.5	49.8	41.4	Not reported	Trails	Not specified Opus/GPT-5.5 starred in source; Fable leads this row. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 49.8 / GPT-5.5 41.4 / GLM-5.2 Not reported	Open evidence
Reasoning / Vendor test / scoreHLE-Full with toolsTrails	K356	Fable 563	Sol58	57.9	52.2	Not reported	Trails	Not specified Opus/GPT-5.5 starred in source; Fable leads this row. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus 57.9 / GPT-5.5 52.2 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / scoreMMMU-ProTrails	K381.6	Fable 581.2	Sol83	78.9	81.2	Not reported	Trails	Official protocol Images precede text; most vision rows averaged three runs. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 78.9 / GPT-5.5 81.2 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / scoreMMMU-Pro with PythonTrails	K383.4	Fable 586.5	Sol84.6	82.7	83.2	Not reported	Trails	Official protocol Most vision rows averaged three runs. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 82.7 / GPT-5.5 83.2 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / scoreCharXiv RQTrails	K384.8	Fable 588.9	Sol84.6	80.5	84.1	Not reported	Trails	Not specified Most vision rows averaged three runs. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 80.5 / GPT-5.5 84.1 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / scoreCharXiv RQ with PythonTrails	K391.3	Fable 593.5	Sol89.1	89.9	89	Not reported	Trails	Not specified Most vision rows averaged three runs. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 89.9 / GPT-5.5 89 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / scoreMathVisionTrails	K394.3	Fable 594.8	Sol95.8	86.7	92.2	Not reported	Trails	Not specified Most vision rows averaged three runs. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 86.7 / GPT-5.5 92.2 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / scoreMathVision with PythonTrails	K397.8	Fable 598.6	Sol97.8	97.1	96.8	Not reported	Trails	Not specified Most vision rows averaged three runs. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 97.1 / GPT-5.5 96.8 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / scoreBabyVision with PythonTrails	K385.7	Fable 590.5	Sol88.9	81.2	83.6	Not reported	Trails	Not specified Most vision rows averaged three runs. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 81.2 / GPT-5.5 83.6 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / pass@5ZeroBench mainTied	K323	Fable 523	Sol17	17	22	Not reported	Tied	Official setting Run five times; K3 ties Fable. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 5 runs Captured: Jul 17, 2026 Other models: Opus 17 / GPT-5.5 22 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / pass@5ZeroBench main with PythonTrails	K341	Fable 546	Sol35	34	41	Not reported	Trails	Official setting Run five times. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 5 runs Captured: Jul 17, 2026 Other models: Opus 34 / GPT-5.5 41 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / scoreWorldVQA ForceAnswerTrails	K351	Fable 556.7	Sol41.8	39.1	38.5	Not reported	Trails	Not specified Most vision rows averaged three runs. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 39.1 / GPT-5.5 38.5 / GLM-5.2 Not reported	Open evidence
Vision / Vendor test / scoreOmniDocBenchLeads	K391.1	Fable 589.8	Sol85.8	87.9	89.4	Not reported	Leads	Not specified K3 leads the published row; most vision rows averaged three runs. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 87.9 / GPT-5.5 89.4 / GLM-5.2 Not reported	Open evidence
Vision / Internal suite / scorePerceptionBenchTrails	K358.5	Fable 557.2	Sol59.7	47.2	55.8	Not reported	Trails	Not specified Moonshot in-house atomic-visual-perception benchmark. Configuration record Config: K3 max effort; temperature 1.0; top-p 1.0 Effort: K3: max Fallback: No K3 fallback reported. Runs: 3 runs Captured: Jul 17, 2026 Other models: Opus 47.2 / GPT-5.5 55.8 / GLM-5.2 Not reported	Open evidence
Independent / Independent / indexArtificial Analysis Intelligence Index v4.1Trails	K357	Fable 560	Sol59	Not reported	Not reported	Not reported	Trails	Kimi API Independent composite; K3 ranked fourth at capture behind Fable and two Sol effort variants. Configuration record Config: Kimi API; provider or leaderboard capture Effort: Captured provider configuration Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 17, 2026 Other models: Opus Not reported / GPT-5.5 Not reported / GLM-5.2 Not reported	Open evidence
Independent / Independent / EloArena WebDev OverallLeads	K31,679 +/-17Preliminary leaderboard result	Fable 51,631	Sol1,618	Not reported	Not reported	1,587	Leads	Arena system Preliminary K3 score with +/-17 interval; page captured July 16. Configuration record Config: Arena system; provider or leaderboard capture Effort: Captured provider configuration Fallback: No K3 fallback reported. Runs: Not reported Captured: Jul 16, 2026 Other models: Opus Not reported / GPT-5.5 Not reported / GLM-5.2 1,587	Open evidence

Cells are not blended. No composite winner score is shown by design.

Twelve-model comparisonSort the supplied field by model, independent intelligence, context, or published price snapshot.

Sort comparison

Twelve-model comparison sortable by model, Artificial Analysis intelligence, context, or price
	Role				Weights	Strongest fit	Biggest caveat
ModelClaude Fable 5	Rolefrontier ceiling	AA snapshot60	Context1M	Price snapshot$10 input / $50 output	WeightsClosed	Strongest fitHardest reasoning and agent work	Biggest caveatHighest cost; fallback/access/retention considerationsAA independent plus official sources
ModelGPT-5.6 Sol	Rolefrontier ceiling	AA snapshot59	Context1.05M	Price snapshot$5 input / $30 output	WeightsClosed	Strongest fitPremium coding; agents; vision; lower token use	Biggest caveatHigher sticker price and product-specific long-context rulesAA independent plus official sources
ModelKimi K3	Rolesubject	AA snapshot57	Context1M	Price snapshot$0.30 cached / $3 input / $15 output	WeightsPromised July 27; not shipped	Strongest fitLong-horizon agents; WebDev; visual coding; research workflows	Biggest caveatVerbose; 62 t/s; preserved-thinking contract; 64+ accelerator guidanceAA independent plus Moonshot official
ModelGLM-5.2	Roleavailable open-weight leader	AA snapshot51	Contextsource-specific	Price snapshotProvider dependent	WeightsAvailable	Strongest fitCurrent highest AA-scored downloadable open weights	Biggest caveatLower AA score than K3; provider varianceAA independent
ModelMiniMax M3	Roleopen-weight cohort	AA snapshot44	Contextsource-specific	Price snapshotProvider dependent	WeightsAvailable	Strongest fitOpen-weight agent/value option	Biggest caveatBelow K3 and GLM-5.2 AA indexAA independent
ModelDeepSeek V4 Pro	Roleopen-weight cohort	AA snapshot44	Contextsource-specific	Price snapshotProvider dependent	WeightsAvailable	Strongest fitOpen-weight reasoning/value option	Biggest caveatMax-effort and provider varianceAA independent
ModelClaude Opus 4.8	Roleprior premium	AA snapshotNot captured	Context1M	Price snapshotOfficial provider pricing	WeightsClosed	Strongest fitMature Anthropic agent baseline	Biggest caveatBehind current Fable generationMoonshot table
ModelGPT-5.5	Roleprior premium	AA snapshotNot captured	Context1.05M	Price snapshotOfficial provider pricing	WeightsClosed	Strongest fitMature Codex-harness baseline	Biggest caveatBehind current Sol generationMoonshot table
ModelKimi K2.7 Code	RoleKimi coding baseline	AA snapshotNot captured	Context256K	Price snapshot$0.19 cached / $0.95 input / $4 output	WeightsAvailable	Strongest fitMature and lower-cost Kimi coding	Biggest caveatLower peak capability and contextMoonshot official
ModelKimi K2.6	RoleKimi general baseline	AA snapshotNot captured	Context256K	Price snapshot$0.55 input / $2.65 output	WeightsAvailable	Strongest fitFast general Kimi mode	Biggest caveatPrior generationKimi product plus prior official research
ModelGemini 3.1 Pro	Rolebuyer context	AA snapshotNot captured	Context1M	Price snapshot$2 input / $12 output	WeightsClosed	Strongest fitBroad multimodal reasoning	Biggest caveatPreview/surface status and not in Moonshot main tableOfficial model docs; recheck
ModelGrok 4.5	Rolebuyer context	AA snapshotNot captured	Context500K	Price snapshot$2 input / $6 output	WeightsClosed	Strongest fitCoding-agent value and live X research	Biggest caveatSmaller context and regional accessExisting July 11 verified packet

Architecture and operating guidanceSee how 16 of 896 experts activate, then review Moonshot's preserved-thinking and permission warnings.

K3 contains 2.8 trillion total parameters, but each token wakes only 16 specialist groups out of 896.

Specialists are selected token by token.
Long inputs are compressed and routed without waking the whole model.
Training emphasizes stable reasoning, vision, and long-agent traces.
Sparse compute still requires storing and serving the full checkpoint.

Sparse activation reduces active compute; it does not make the full model laptop-friendly.

16 active experts896 total

Preserved thinking

Mid-session model switch

Strong permission boundaries

Editorial risk reading

Reasoning history remains available across turns, matching Moonshot's recommended operating pattern.
The same model remains in the loop, reducing state mismatch risk.
Tool permissions are constrained, limiting the blast radius of excessive proactivity.

This is why the benchmark table carries a harness column. It is guidance, not a numerical safety score.

Launch demos and Matt's testing boundaryTen vendor showcases remain reproduction targets; Matt's planned ten-task quality test returned no model output because access was blocked.

GPU kernel optimization

Vendor Test

Moonshot reports up to 24-hour autonomous optimization across four GPU-kernel tasks.

Still needed: Publish trajectories, budgets, hardware, tolerances, and same-sandbox independent runs.

MiniTriton compiler

Vendor Test

Moonshot reports K3 building a Triton-like compiler with MLIR, optimization passes, PTX generation, and nanoGPT training.

Still needed: Release the repository, tests, unsupported cases, and an independent performance audit.

Procedural 3D open world

Vendor Test

Moonshot demonstrates a browser-based procedural 3D open world with vision in the loop.

Still needed: Publish the full prompt, code, interventions, asset rights, token budget, and repeat runs.

48-hour chip-design proof of concept

Vendor Test

Moonshot reports a 48-hour chip-design proof of concept using open-source EDA tools.

Still needed: Release RTL, netlists, verification artifacts, PPA methodology, and independent physical validation.

I-Love-Q astrophysics research

Vendor Test

Moonshot demonstrates an astrophysics workflow spanning papers, equations of state, Python, and an interactive dashboard.

Still needed: Publish sources, code, environment, calculation checks, and a domain-expert review.

42-year AI ASIC research site

Vendor Test

Moonshot reports a 42-year ASIC-industry research site built through many recursive iterations and source pulls.

Still needed: Publish elapsed time, cost, source-quality audit, citation checks, and a human-edit log.

Fusion and gravitational-wave research

Vendor Test

Moonshot reports fusion and gravitational-wave research produced with concurrent agents.

Still needed: Publish the complete methodology, source set, agent traces, and expert validation.

Slides, Docs, and Sheets artifacts

Vendor Test

Moonshot shows editable presentations, documents, spreadsheets, heatmaps, and annual-report artifacts.

Still needed: Audit formula accuracy, export fidelity, edit burden, accessibility, and repeatability.

Widgets and Dashboard

Vendor Test

Moonshot presents persistent Widgets and Dashboard features connected to data and plugins.

Still needed: Document connector coverage, refresh behavior, permissions, data handling, and failure states.

Motion graphics and 56-clip teaser edit

Vendor Test

Moonshot reports motion graphics and a teaser edit assembled from 56 clips with revisions and beat sync.

Still needed: Publish the project file, timeline, human direction, media-rights audit, and a repeat run.

EditorialStatus: not run / access blocked

K3 Max was selectable anonymously, but the first prompt opened a login modal and returned no output. The page does not claim an independent Matt Farmer model-quality result.

T01frontend generationnot run access blocked
T02visual to codenot run access blocked
T03multi file debuggingnot run access blocked
T04long horizon recoverynot run access blocked
T05cited researchnot run access blocked
T06long context retrievalnot run access blocked
T07office spreadsheetnot run access blocked
T08instruction boundariesnot run access blocked
T09multilingualnot run access blocked
T10multimodalnot run access blocked

Economics, privacy, reaction, and methodologyInspect all six cost presets, the standard-API privacy checklist, community questions, release facts, corrections, and research changelog.

Six supplied cost presets

Six Kimi K3 cost worksheet presets
Scenario	Cached input	Cache-miss input	Output	Turns	Total
Short chat	0	2,000	1,000	1	$0.0210
Coding turn with warm repository	90,000	10,000	20,000	1	$0.3570
Research task	375,000	125,000	80,000	1	$1.6875
Full-context miss	0	1,000,000	100,000	1	$4.5000
Ten-turn agent loop	80,000	20,000	10,000	10	$2.3400
Artificial Analysis blended MTok	700,000	200,000	100,000	1	$2.3100

Standard API privacy questions

Zero retention

Not verified as a standard default

No training on inputs

Not verified as a standard default

Data residency

Not established in this review

Sensitive workloads

Require security, legal, and procurement review

Read the current terms

Community questions worth testing

Frontier watchers

Frontier shock

Many users interpret the scores as the first open-weight release to approach the closed frontier.

Frontend leadership

Arena listed K3 first on WebDev Overall at 1679 with a preliminary label.

Price pressure

Some praise half-Sol sticker pricing; others call $3/$15 expensive for a Chinese open model.

US-China framing

Launch coverage and forums frame K3 as evidence that the capability gap is shrinking.

Working developers

Token hunger

Artificial Analysis measured 130M output tokens across its index; users warn sticker price may understate finished-task cost.

Benchmaxxing concern

Users want real-world reproduction before replacing established subscriptions.

Conflicting real-world quality

Some prompts impress; others reportedly trail Sol/Fable or even K2.6 on factuality and task quality.

Persistence

A Kimi user reported a four-hour run that eventually completed; it signals persistence and potentially extreme time/token cost.

Excessive proactivity

Moonshot warns K3 may make unexpected decisions on ambiguous tasks.

Thinking-history sensitivity

Moonshot warns that missing preserved reasoning history or switching models mid-session can destabilize quality.

Local-model operators

Self-hosting excitement

Users value sovereignty, customization, and reduced vendor lock-in.

Self-hosting reality

Local-model users joke that a 2.8T release is open but not practically runnable at home.

Release facts and corrections

Launch: July 16, 2026; Official
API model ID: kimi-k3; Official
Parameters: 2.8T total (vendor-stated); Official
Expert routing: 16 active of 896; Official
Context: 1M tokens; Official
Input: Text, images, and documented video-file input; Official
API price: $0.30 cached / $3 miss / $15 output per 1M; Official
Reasoning effort: API: max only; Kimi Code docs: low / high / max; Official
Weights: Promised by July 27; not shipped at cutoff; Pending

Correction / Jul 15

K3 had already launched through a limited recharge campaign.

Verified: The campaign page was user-reported and unverified. Moonshot launched K3 publicly on July 16.

Correction / Jul 15

K3 had about 2.5 trillion parameters.

Verified: Moonshot's launch specification says 2.8 trillion total parameters.

Correction / Jul 15

The mystery Arena model Kivine was confirmed as K3.

Verified: Kivine was community attribution, not official identification. The released leaderboard entry is kimi-k3.

Correction / Jul 15 → 17

A 1M context window meant the full model would immediately be open source and locally downloadable.

Verified: The 1M context was confirmed, but weights, license, and the technical report remained pending until the promised July 27 checkpoint.

Method and changelog

Official specifications and prices come from Moonshot. The independent headline numbers come from Artificial Analysis and Arena. Vendor tests retain their original harness and comparability notes. Community reports identify questions, not conclusions. No unlike scores are blended into a synthetic ranking.

v0.1 / Jul 15, 2026Pre-launch rumor-control packet kept the model ID, price, size, context, and access claims unverified.
v1.0 / Jul 17, 2026Post-launch packet added official specifications, pricing, benchmark rows, independent measurements, access status, terms, and the July 27 watchlist.
v1.0 check / Jul 17, 2026Publication check confirmed the launch, pricing, Arena Preliminary result, Artificial Analysis snapshot, terms, and that K3 weights/report were still absent from Moonshot's public organizations.

Complete source indexOfficial launch, API, pricing, terms, repositories, independent measurements, and dated community evidence.

1. Moonshot: Kimi K3 launch
Specifications, benchmark table, availability, architecture, deployment guidance, and limitations
2. Kimi K3 API quickstart
Model ID, context, API behavior, max effort, multimodal input, caching, and limits
3. Kimi K3 API pricing
Official pricing and billing notes
4. Kimi API rate limits
Recharge and rate-limit terms
5. Kimi OpenPlatform terms
Customer-content use and enterprise agreement language
6. Kimi Code model configuration
Kimi Code plan access, context, and low/high/max effort listing
7. Moonshot AI on Hugging Face
Negative checkpoint check at the July 17 cutoff
8. Moonshot AI on GitHub
Negative technical-report and repository check at the July 17 cutoff
9. Artificial Analysis: Kimi K3
Independent intelligence, speed, latency, pricing, and token-use snapshot
10. Arena WebDev leaderboard
Independent Preliminary WebDev Overall snapshot
11. Frontier watchers: Frontier shock
Anecdotal community evidence checked 2026-07-17
12. Frontier watchers: Frontend leadership
Independent evidence checked 2026-07-17
13. Local-model operators: Self-hosting excitement
Anecdotal community evidence checked 2026-07-17
14. Working developers: Token hunger
Independent evidence checked 2026-07-17
15. Working developers: Benchmaxxing concern
Anecdotal community evidence checked 2026-07-17
16. Working developers: Conflicting real-world quality
Anecdotal community evidence checked 2026-07-17
17. Working developers: Persistence
Anecdotal community evidence checked 2026-07-17
18. Local-model operators: Self-hosting reality
Anecdotal community evidence checked 2026-07-17
19. Frontier watchers: US-China framing
Anecdotal community evidence checked 2026-07-17
20. Working developers: Thinking-history sensitivity
Official evidence checked 2026-07-17

08 / Questions

What people actually want to know

These are the short answers I would give a smart friend before they tried K3, compared it with another model, or planned around the promised weights.

What is Kimi K3?

Kimi K3 is Moonshot AI's 2.8-trillion-parameter sparse mixture-of-experts model with native visual understanding, a 1-million-token context window, agentic products, and an API.

When did Kimi K3 launch?

Moonshot launched Kimi K3 on July 16, 2026. This evidence snapshot was last verified on July 17, 2026.

Is Kimi K3 better than GPT-5.6 Sol or Claude Fable 5?

Not overall based on the July 17 evidence. K3 led selected vendor rows and a Preliminary Arena WebDev snapshot, while Artificial Analysis placed Claude Fable 5 and GPT-5.6 Sol ahead on its independent composite.

How much does the Kimi K3 API cost?

Moonshot lists $0.30 per million cached input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens. Tool fees, retries, and taxes can add to the completed-job cost.

Is Kimi K3 open source or open weight?

Neither status was verifiable at the July 17 cutoff. Moonshot promised full weights and a technical report by July 27, but the checkpoint and license were not yet available. Open weight and open source are not interchangeable.

Eight more technical and deployment questions

Can I download Kimi K3 now?

No verified full Kimi K3 checkpoint was available on July 17. Use Moonshot's official Hugging Face and GitHub organizations to verify any later release.

Can Kimi K3 run on a consumer PC or Mac?

Not the full 2.8T model in any practical sense. Moonshot recommends a supernode with 64 or more accelerators for deployment; that is vendor guidance, not an independently measured minimum.

Why does preserved thinking history matter?

Moonshot says K3 was trained with preserved thinking history and can become unstable when a harness omits the complete historical assistant message or switches models mid-session.

What does Moonshot mean by excessive proactivity?

Moonshot warns that K3 may make unexpected decisions when a task has minor problems or ambiguous intent. Production agents should use explicit permissions, budgets, stop conditions, and approval boundaries.

What did independent Kimi K3 benchmarks find?

At the July 17 capture, Artificial Analysis scored K3 at 57 with 62 output tokens per second and 130 million output tokens across its Intelligence Index run. Arena listed K3 first on WebDev Overall at 1679 +/-17, marked Preliminary, on July 16.

Why can a cheaper per-token model cost more per task?

Completed-job cost depends on output volume, turns, retries, caching, tool fees, latency, and cleanup. K3's $15-per-million output rate can dominate the bill when the model reasons or retries at length.

Does Moonshot use API data for model improvement?

Moonshot's captured API terms say customer content may be used to provide, maintain, develop, support, improve, secure, and enforce the service unless a separate written enterprise arrangement restricts that use. Review the current terms before sending sensitive data.

What changes on July 27, 2026?

Moonshot promised full model weights by July 27. A responsible update must verify the checkpoint, license, technical report, model configuration, vLLM support, third-party hosts, and independent self-host evidence before changing the page's pending status.

The practical next step

Try K3 on a real job—not a victory-lap prompt

Give it a meaningful coding, research, document, or visual task. Track output volume and retries. Then compare the finished result, time, and cost with the model you already use.

Try Kimi K3

Work with Matt

Need help choosing an AI model stack?

Book a strategy hour with Matt to compare quality, cost, privacy, and workflow fit in plain English—and leave with a model plan your team can actually use.

Work with Matt

Kimi K3 explained: China's new AI heavyweight.

A Chinese AI lab just moved closer to the frontier

How close is K3 to GPT-5.6 and Claude?

Independent intelligence index

Preliminary Arena WebDev score

K3 is cheaper—but output still matters

Published API price per million tokens

Quick chat

Coding with a familiar project

Long research task

What can ordinary people do with K3?

Build websites and software

Research very long documents

Create office work

Understand images

What “open weight” actually means

Open weight

Open source

Read this before you choose K3

Close is not the same as first

Cheap tokens can become an expensive job

Hosted API privacy needs review

The open-weight story is still a promise

The technical evidence, if you want it

GPU kernel optimization

MiniTriton compiler

Procedural 3D open world

48-hour chip-design proof of concept

I-Love-Q astrophysics research

42-year AI ASIC research site

Fusion and gravitational-wave research

Slides, Docs, and Sheets artifacts

Widgets and Dashboard

Motion graphics and 56-clip teaser edit

Six supplied cost presets

Standard API privacy questions

Community questions worth testing

Frontier watchers

Working developers

Local-model operators

Release facts and corrections

Method and changelog

What people actually want to know

What is Kimi K3?

When did Kimi K3 launch?

Is Kimi K3 better than GPT-5.6 Sol or Claude Fable 5?

How much does the Kimi K3 API cost?

Is Kimi K3 open source or open weight?

Can I download Kimi K3 now?

Can Kimi K3 run on a consumer PC or Mac?

Why does preserved thinking history matter?

What does Moonshot mean by excessive proactivity?

What did independent Kimi K3 benchmarks find?

Why can a cheaper per-token model cost more per task?

Does Moonshot use API data for model improvement?

What changes on July 27, 2026?

Try K3 on a real job—not a victory-lap prompt

Need help choosing an AI model stack?