Inference
EN
Moonshot AI

Kimi K2.5

Kimi K2.5 is a standard model from Moonshot AI. 55 benchmarks count toward its index, in 8 categories.

  • Open weights
  • Standard
  • Input: text · image
  • Released 2026-01
  • Retires 18 Nov 2026
RankedEnough results for an index.
IndexAll categories, out of 100.52.2±2.2Median 51.0
CoverageIndex weight that has results.100%
SpeedOutput tokens per second.36/sMedian 67/s
First tokenSeconds to the first output token.0.9 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$0.532Median $0.60
Output / 1MUS dollars per 1M output tokens.$3.32Median $2.55
ContextMaximum tokens in one request.262KMedian 262K
EloLMArena rating and rank.—Median 1422

Details

From OpenRouter
API IDThe model ID on OpenRouter.
moonshotai/kimi-k2.5
Max outputMaximum output tokens per request.
66K
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.
$0.209
WeightsLink to the published weights.
Hugging Face (opens in a new tab) ↗
RetiresDate OpenRouter removes the model.
18 Nov 2026

Where to run

APIs that serve this model.
6 providers
ProviderTypeMaker, cloud, host or router.InputOutputContext
Deep InfraDeprecatedHost$0.45$2.25262K
SiliconFlowHost$0.45$2.25262K
Amazon BedrockCloud$0.60$3.00262K
Microsoft AzureCloud$0.60$3.00262K
Novita AIHost$0.60$3.00262K
OpenRouterRouter$0.532$3.33262K

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

Run locally

1026.88B parameters
16-bit
2157.4 GB
8-bit
1146.6 GB
4-bit
654.7 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.16 counted49.4132 / 218
CodingCode writing and repair.6 counted53.9107 / 209
ReasoningLogic problems and puzzles.5 counted51.6104 / 227
MultimodalTasks with images and text.8 counted52.464 / 84
KnowledgeFacts and expert knowledge.8 counted53.7106 / 268
MultilingualTasks in many languages.2 counted52.75 / 9
InstructionTasks with strict prompt rules.2 counted53.529 / 63
MathMath problems.8 counted53.871 / 168

Results

55 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
C-EvalKnowledge94.0%1 / 12
τ²-Bench Tool-Agent-User EvaluationAgentic95.9%13 / 149
Massive Multitask Language Understanding ProfessionalKnowledge87.1%7 / 75
American Invitational Mathematics Examination 2025Math96.1%6 / 53
Instruction-Following EvalInstruction93.9%10 / 70
Massive Multi-discipline Multimodal UnderstandingMultimodal84.3%5 / 28
SWE-bench Verified · SWE-bench/experimentsCoding70.8%13 / 65
MMLU-ReduxKnowledge94.5%4 / 19
AIME 2026Math95.8%8 / 34
Scientific Code BenchmarkCoding48.7%9 / 35
Artificial Analysis GPQA DiamondKnowledge87.9%53 / 199
Harvard-MIT Mathematics Tournament February 2025Math95.4%4 / 15
OTIS Mock AIME 2024-2025Math92.2% ±3.954 / 200
GPQA diamond · Epoch AIKnowledge87.6% ±1.961 / 217
Graduate-Level Google-Proof Q&AKnowledge87.6%41 / 141
Artificial Analysis IFBenchInstruction70.2%44 / 146
FrontierMath-2025-02-28-PrivateMath27.9% ±2.621 / 69
LiveCodeBench v6Coding85.0%19 / 59
Artificial Analysis Long Context ReasoningReasoning78.0%70 / 217
MMLU-Pro first-party comparison snapshotKnowledge87.1%2 / 6
Software Engineering Benchmark VerifiedCoding76.8%38 / 114
Artificial Analysis Humanity's Last ExamKnowledge30.7%74 / 222
Berkeley Function Calling Leaderboard v4Agentic68.3%10 / 26
Artificial Analysis Omniscience AccuracyKnowledge35.2%65 / 164
CountBenchMultimodal94.1%4 / 10
τ³-Bench Tool-Agent-User EvaluationAgentic65.7%13 / 32
Artificial Analysis MMMU-ProMultimodal75.4%49 / 113
Massive Multi-discipline Multimodal Understanding ProMultimodal78.5%23 / 53
LongBench v2Reasoning61.0%11 / 25
Critical Physics TasksReasoning3.1%95 / 214
SuperGPQA: Scaling LLM Evaluation Across 285 Graduate DisciplinesKnowledge69.2%10 / 22
MMLU-ProXMultilingual82.3%8 / 17
FrontierMath-Tier-4-2025-07-01-PrivateMath4.2% ±2.926 / 55
Harvard-MIT Mathematics Tournament February 2026Math87.1%13 / 27
INCLUDEMultilingual83.3%6 / 12
SimpleVQAMultimodal71.2%9 / 18
ARC-AGI-1 (semi-private)Reasoning65.3%63 / 126
QwenClawBenchAgentic54.3%5 / 10
ARC-AGI-2 (semi-private)Reasoning11.8%66 / 131
SWE-Bench verified · Epoch AICoding73.8% ±2.017 / 32
Gert Labs Composite Game BenchmarkAgentic45.9%29 / 52
GPQA Diamond · BenchLMKnowledge87.6%47 / 84
React Native EvalsCoding77.2%9 / 16
Harvard-MIT Mathematics Tournament November 2025Math91.1%7 / 12
RealWorldQAMultimodal81.0%11 / 18
Humanity's Last ExamKnowledge30.1%48 / 77
GDPval-AA normalizedAgentic17.0%83 / 133
SimpleQA VerifiedKnowledge34.3% ±1.559 / 91
DeepSearchQAAgentic77.1%16 / 24
Terminal-Bench 2.0 · BenchLMAgentic50.8%28 / 41
RefCOCO averageMultimodal87.8%7 / 10
MathVisionMultimodal84.2%17 / 24
WideResearchAgentic72.7%12 / 16
OSWorld-VerifiedAgentic63.3%29 / 37
SWE-bench Verified (mini-swe-agent-v2)Coding70.8%4 / 5
CharXiv ReasoningMultimodal77.5%40 / 49
APEX-Agents-AAAgentic11.5%25 / 30
BrowseCompAgentic60.6%53 / 63
SWE-bench ProCoding50.7%70 / 83
Claw-EvalAgentic52.3%37 / 42
NL2RepoCoding32.0%34 / 36
ToolathlonAgentic27.8%21 / 22
JobBenchAgentic8.7%35 / 36
MCP AtlasAgentic29.5%52 / 52
ARC-AGI-2 (public eval)Reasoning12.1%58 / 122
ARC-AGI-1 (public eval)Reasoning73.1%58 / 116
SWE-bench Multilingual · SWE-bench/experimentsCoding67.3%7 / 14
SWE-bench Multilingual · swebench.comMultilingual67.3%7 / 12
AIME25 first-party comparison snapshotMath96.3%—
AI2D test splitMultimodal90.8%—
OmniDocBench 1.5Multimodal88.8%—
Video-MMEMultimodal87.4%—
Video-MME with subtitleMultimodal87.4%—
VideoMMMUMultimodal86.6%—
MLVU mean averageMultimodal85.0%—
We-MathMultimodal84.7%—
DynaMathMultimodal84.4%—
Video-MME without subtitleMultimodal83.2%—
MMAnswerBenchMath81.8%—
Multimodal Multi-disciplinary Video UnderstandingMultimodal80.4%—
CC-OCRMultimodal79.7%—
MAXIFEMultilingual72.8%—
MCP-TasksAgentic59.1%—
SWE-RebenchCoding58.5%—
MMLongBench-DocMultimodal58.5%—
NOVA-63Multilingual56.0%—
PolyMathMultilingual43.1%—
DeepPlanningAgentic14.4%—
ResearchClawBenchAgentic14.0%—

Sources

BenchLMCC BY-NC 4.0OpenRouterTerms restrict reuseEpoch AICC BYSWE-bench teamNo licence statedARC Prize FoundationTerms restrict reusemodels.devMIT

History

5 releases
Index53.3 → 52.2
23 Oct 202612 Nov 2026
Output / 1M$2.50 → $3.33
23 Oct 202612 Nov 2026

Questions

How much does Kimi K2.5 cost?
$0.532 per 1M input tokens, $3.32 per 1M output tokens.
How good is Kimi K2.5?
Its index is 52.2 out of 100. That is place 127 of 283 ranked models.
What is Kimi K2.5 best at?
Knowledge: 53.7 out of 100, place 106 of 268.
How fast is Kimi K2.5?
36 output tokens per second, 0.9 s to the first token.
What is the context window of Kimi K2.5?
262K tokens.
Is Kimi K2.5 open source?
Yes. Its weights are published.
When was Kimi K2.5 released?
January 2026.

Badge

Kimi K2.5 on Inference360: place 127 of 283, index 52.2
HTML
<a href="https://inference360.ai/models/kimi-k2-5"><img src="https://inference360.ai/badges/kimi-k2-5.svg" alt="Kimi K2.5 on Inference360: place 127 of 283, index 52.2" height="20"></a>
Markdown
[![Kimi K2.5 on Inference360: place 127 of 283, index 52.2](https://inference360.ai/badges/kimi-k2-5.svg)](https://inference360.ai/models/kimi-k2-5)

Compare with

Ternary Bonsai 2 27B52.2Hy3 Preview52.3Ripple-65B Preview52.3Solar Pro 452.1Kimi K2.5 (Reasoning)53.0

Lower price

dots3-note Preview63.4 · FreeApodex 1.1 Mini55.7 · FreeNemotron 3 Ultra50.8 · Free

More from Moonshot AI

All models
Kimi K2.5 (Reasoning)53.0Kimi K2 Thinking48.2Kimi K2.658.2Kimi LatestUnrankedKimi K2.7 Code57.7