Inference
EN
Moonshot AI

Kimi K3

Kimi K3 is a reasoning model from Moonshot AI. 73 benchmarks count toward its index, in 7 categories.

  • Open weights
  • Reasoning
  • Input: text · image · video
  • Released 2026-07
RankedEnough results for an index.
IndexAll categories, out of 100.66.9±2.4Median 51.0
CoverageIndex weight that has results.95%
SpeedOutput tokens per second.44/sMedian 67/s
First tokenSeconds to the first output token.3.3 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$0.192batch $2.28Median $0.60
Output / 1MUS dollars per 1M output tokens.$16.00batch $11.40 Batch price per 1M output tokens.Median $2.55
ContextMaximum tokens in one request.1.05MMedian 262K
EloLMArena rating and rank.1474±4 #27Median 1422

Details

From OpenRouter
API IDThe model ID on OpenRouter.
moonshotai/kimi-k3
Max outputMaximum output tokens per request.
944K
Knowledge cutoffLast date of training data.
—
ToolsTool calls on OpenRouter.
Yes
JSON outputOutput follows a given JSON schema.
Yes
Cache read / 1MUS dollars per 1M cached tokens.
$0.182
WeightsLink to the published weights.
Hugging Face (opens in a new tab) ↗
RetiresDate OpenRouter removes the model.
—

Where to run

APIs that serve this model.
17 providers
ProviderTypeMaker, cloud, host or router.InputOutputContext
ChutesHost$1.85$9.501.05M
Fireworks AIHost$2.10$10.501.05M
SiliconFlowHost$2.70$13.501.05M
Deep InfraHost$2.85$14.251.05M
Alibaba CloudHost$3.00$15.001.05M
Amazon BedrockCloud$3.00$15.001.05M
Arcee AIHost$3.00$15.001.05M
BasetenHost$3.00$15.001.05M
DigitalOceanCloud$3.00$15.001.05M
ModalHost$3.00$15.001.05M
Moonshot AIMaker$3.00$15.001.05M
NebiusHost$3.00$15.001.02M
Novita AIHost$3.00$15.001.05M
Together AIHost$3.00$15.001.05M
OpenRouterRouter$0.192$16.001.05M
Berget AIHost$3.36$16.82328K
TinfoilHost$4.00$20.00262K

US dollars per 1M tokens, list price. From models.dev (opens in a new tab) (MIT).

Run locally

2779.93B parameters
16-bit
5838.9 GB
8-bit
3102.4 GB
4-bit
1770.6 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy needs more than 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.27 counted65.941 / 218
CodingCode writing and repair.19 counted66.235 / 209
ReasoningLogic problems and puzzles.8 counted67.735 / 227
MultimodalTasks with images and text.7 counted65.624 / 84
KnowledgeFacts and expert knowledge.6 counted66.338 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.1 counted71.0—
MathMath problems.5 counted66.627 / 168

Results

73 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
Artificial Analysis Long Context ReasoningReasoning88.7%1 / 217
Artificial Analysis GPQA DiamondKnowledge93.5%8 / 199
Graduate-Level Google-Proof Q&AKnowledge93.5%10 / 141
DeepSearchQAAgentic95.0%2 / 24
MathVisionMultimodal94.3%2 / 24
GPQA Diamond · Vals AIKnowledge92.9% ±1.312 / 135
SWE-benchCoding93.4% ±1.18 / 87
Artificial Analysis SciCodeCoding59.5%11 / 114
Artificial Analysis Humanity's Last ExamKnowledge46.9%22 / 222
APEX-Agents-AAAgentic41.3%3 / 30
GPQA diamond · Epoch AIKnowledge93.1% ±1.522 / 217
CharXiv ReasoningMultimodal91.3%5 / 49
Terminal-Bench 2.1 · BenchLMAgentic88.3%9 / 86
Critical Physics TasksReasoning23.4%23 / 214
MMMU ProMultimodal88.2% ±0.810 / 92
BrowseCompAgentic91.2%7 / 63
GPQA Diamond · BenchLMKnowledge93.5%10 / 84
Terminal-Bench 2.1 · Vals AIAgentic80.9% ±0.69 / 75
LiveBench ReasoningReasoning90.7%10 / 80
LiveCodeBenchCoding87.2% ±1.018 / 142
Massive Multi-discipline Multimodal Understanding ProMultimodal81.6%7 / 53
MCP AtlasAgentic84.2%7 / 52
Vibe Code Bench v1.1Coding85.0% ±2.717 / 123
LiveBench LanguageKnowledge85.5%10 / 66
LiveBench Agentic CodingAgentic62.2%10 / 66
Artificial Analysis MMMU-ProMultimodal80.5%18 / 113
MMLU ProKnowledge88.0% ±0.322 / 135
ARC-AGI-1 (semi-private)Reasoning94.5%21 / 126
Humanity's Last ExamKnowledge56.0%13 / 77
OTIS Mock AIME 2024-2025Math97.2% ±1.134 / 200
Artificial Analysis Omniscience AccuracyKnowledge47.6%28 / 164
GDPval-AA normalizedAgentic51.6%26 / 133
LiveBench CodingCoding81.4%13 / 66
ProofBench v1.1Math87.0% ±3.413 / 54
LiveBench Data AnalysisReasoning78.7%22 / 80
FrontierMath-Tiers-1-3-v2-PrivateMath72.2% ±2.727 / 90
ARC-AGI-2 (semi-private)Reasoning60.4%40 / 131
Artificial Analysis Tau3-BankingAgentic46.0%7 / 22
SimpleQA VerifiedKnowledge50.6% ±1.629 / 91
MazeBench: visual spatial reasoning in a 3D open worldAgentic1.0%16 / 48
LiveBench Instruction FollowingInstruction71.4%23 / 66
CharXiv Reasoning without toolsMultimodal84.8%6 / 17
LiveBench Long AgenticAgentic42.8%18 / 51
Humanity's Last Exam without toolsKnowledge43.5%19 / 53
MirrorCodeCoding54.4% ±11.414 / 39
ProgramBench: Can Language Models Rebuild Programs From Scratch?Coding77.8%7 / 19
Artificial Analysis ITBench-AAAgentic47.7%9 / 24
FrontierMath-Tier-4-v2-PrivateMath39.0% ±7.728 / 72
LiveBench ResearchAgentic79.0%21 / 52
Agent Arena task outcomeAgentic3.422 / 54
JobBenchAgentic52.9%15 / 36
Agent Arena command recoveryAgentic4.421 / 50
Artificial Analysis CyberGym-E2E-AAAgenticAbove fitFar above this model's index.58.0%8 / 19
Artificial Analysis Cyber IndexAgentic41.3%8 / 19
LiveBench DesignAgentic77.2%22 / 52
Artificial Analysis Terminal-Bench v2.1Agentic85.0%9 / 21
τ²-bench BankingAgentic37.1%12 / 28
Hermes Index: agentic capability and cost per task inside Hermes AgentAgentic39.2%7 / 16
Mystery Game PuzzlesReasoning26.0% ±4.435 / 80
OpenHarmony Bench v1.0Coding57.3%6 / 13
Chess PuzzlesReasoning36.0% ±4.842 / 90
DeepSWEAgentic67.5%28 / 59
ApprenticeBench: end-to-end computer use, continual learning, and long-horizon agency on a real accounts-payable jobAgentic18.0%9 / 18
Agent Arena steerabilityAgentic−0.828 / 54
APEX-AgentsAgentic37.6%7 / 13
OfficeQA ProMultimodal63.3%9 / 16
Toolathlon-VerifiedAgentic73.2%13 / 23
Vibe Code Bench 1-100Coding18.2% ±3.915 / 26
IOICoding48.9% ±9.826 / 45
PostTrainBench v1.1Coding32.0%9 / 15
Terminal-Bench 4.0Agentic17.2% ±0.537 / 61
Artificial Analysis CWE-Bench-AAAgentic53.3%12 / 19
FrontierSWE v2Coding25.9%14 / 22
Furniture AssemblyReasoning34.2% ±6.124 / 37
cursorBench32Coding60.8%12 / 18
ExtractBench: schema-guided enterprise document extraction (LlamaIndex leaderboard)Multimodal83.2%18 / 26
ZeroBenchMultimodal23.0%7 / 10
Code MigrationCoding16.1% ±4.155 / 77
LiveBench MathematicsMath84.4%52 / 66
Bug Hunt BenchCoding21.0%20 / 25
Artificial Analysis DeepsecBench-AAAgentic12.5%17 / 21
ScarfBench: Cross-Framework Application Migration in Enterprise JavaCoding61.3%9 / 11
AutomationBenchAgentic30.8%27 / 32
LiveBench Browser UseAgentic16.0%42 / 49
Artificial Analysis AnalystAgentAgentic38.8%23 / 26
FrontierCode 2.0Coding17.7%24 / 27
LiveBench Coding 2.0Coding59.3%48 / 52
VulcanBench v3Coding73.7%14 / 14
CorpFin v2Finance71.6% ±0.93 / 131
TaxEval v2Finance75.7% ±0.812 / 143
LegalBenchLegal86.2% ±0.415 / 149
MedScribeHealthcare88.0% ±2.015 / 109
CyberBench v1.1Other75.2% ±5.68 / 48
ProgramBenchCoding2.0% ±1.011 / 65
Legal Research BenchLegal46.2% ±3.519 / 90
MedCodeHealthcare49.4% ±2.223 / 106
Tax Agent BenchFinance33.4% ±3.419 / 82
Public Benefits Bench v1.1Public benefits68.2% ±1.212 / 50
EMBFinance66.7% ±2.921 / 75
ARC-AGI-1 (public eval)Reasoning94.9%34 / 116
ARC-AGI-2 (public eval)Reasoning61.7%37 / 122
BioMysteryBenchScience72.6% ±1.610 / 30
MortgageTaxFinance66.3% ±0.935 / 97
SAGEEducation86.8% ±1.912 / 29
HLAB-Vals-1.1Legal12.9% ±2.815 / 35
Finance Agent (v2)Finance53.1% ±0.435 / 79
FrontierMath-ErdosMath0.0%8 / 15
Time Horizon Index: KSPOther10.5% ±5.49 / 13
Terminal-Bench-ScienceScience1.4% ±1.437 / 44
MathVision with PythonMultimodal97.8%—
Artificial Analysis Harvey LAB-AA v1.0Agentic94.6%—
Refusal of harmful cyber requests (JailbreakBench, StrongREJECT, AgentHarm)Agentic94.0%—
OmniDocBenchMultimodal91.1%—
b3: Backbone Breaker BenchmarkAgentic88.0%—
BabyVision with PythonMultimodal85.7%—
MMMU-Pro with PythonMultimodal83.4%—
FrontierSWECoding81.2%—
Finch (FinWorkBench): Finance & Accounting across Spreadsheet-Centric Enterprise WorkflowsAgentic77.3%—
DECK-Bench (Internal)Agentic73.5%—
Kimi Code Bench v2Coding72.9%—
Hermes Bench: 150 agent tasks graded on the files and state the agent leaves behindAgentic71.5%—
PerceptionBench (Internal)Multimodal58.5%—
WorldVQA ForceAnswerMultimodal51.0%—
MLS-Bench LiteCoding48.3%—
Artificial Analysis EnterpriseOps-GymAgentic45.3%—
SWE-MarathonCoding42.0%—
ZeroBench_main with PythonMultimodal41.0%—
PostTrain BenchCoding36.6%—
SpreadsheetBench 2Agentic34.8%—

Sources

BenchLMCC BY-NC 4.0OpenRouterTerms restrict reuseEpoch AICC BYARC Prize FoundationTerms restrict reuseVals AINo licence statedLiveBenchNo licence statedSierra ResearchMITLMArenaCC BY 4.0models.devMIT

History

5 releases
Index67.7 → 66.9
23 Oct 202612 Nov 2026
Output / 1M$11.94 → $16.00
23 Oct 202612 Nov 2026

Questions

How much does Kimi K3 cost?
$0.192 per 1M input tokens, $16.00 per 1M output tokens.
How good is Kimi K3?
Its index is 66.9 out of 100. That is place 34 of 283 ranked models.
What is Kimi K3 best at?
Knowledge: 66.3 out of 100, place 38 of 268.
How fast is Kimi K3?
44 output tokens per second, 3.3 s to the first token.
What is the context window of Kimi K3?
1.05M tokens.
Is Kimi K3 open source?
Yes. Its weights are published.
When was Kimi K3 released?
July 2026.

Badge

Kimi K3 on Inference360: place 34 of 283, index 66.9
HTML
<a href="https://inference360.ai/models/kimi-k3"><img src="https://inference360.ai/badges/kimi-k3.svg" alt="Kimi K3 on Inference360: place 34 of 283, index 66.9" height="20"></a>
Markdown
[![Kimi K3 on Inference360: place 34 of 283, index 66.9](https://inference360.ai/badges/kimi-k3.svg)](https://inference360.ai/models/kimi-k3)

Compare with

Ling 3.1 Flash67.2Hy4 preview66.4Inkling 2 Small66.4Claude Opus 4.867.7Kimi K3.1 Pro75.7

Lower price

Ling 3.1 Flash67.2 · $0.135/1MMiMo-V2.6-Flash64.3 · $0.28/1MDeepSeek V4.1 Flash 101768.5 · $0.40/1M

More from Moonshot AI

All models
Kimi K2.7 Code57.7Kimi LatestUnrankedKimi K2.658.2Kimi K3.1 Pro75.7Kimi K2.5 (Reasoning)53.0