Kimi K2 Thinking

Kimi K2 Thinking places in the top three on 3 benchmarks.

Moonshot AI Released 6 Nov 2025 Family Kimi K2 The lab's model card

Benchmarks with a score

Ranked first, out of five or more models

In the top three

Moonshot AI Nov 2025

Open weights Flagship

145.9 #88 of 271

Epoch Capabilities Index

Where it ranks

Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.

Benchmark	Score	Rank	Source
MMLU-Redux	94.4%	#5 of 38	Model card
MMLU-Pro	84.6%	#35 of 151	Model card
FRAMES	87%	#2 of 8	Model card
Creative Writing v3	73.3	#36 of 120	Official leaderboard
Humanity's Last Exam	44.9%	#50 of 147	Model card
Longform Creative Writing (EQ-Bench)	69.0	#43 of 119	Official leaderboard
Seal-0	56.3%	#2 of 5	Model card
CL-bench	17.6%	#14 of 24	Official leaderboard
MultiChallenge	55.4%	#24 of 35	Official leaderboard
SkateBench v2	43.9%	#16 of 21	Official leaderboard
Longform Writing	73.8%	#1 of 1	Model card

Benchmark	Score	Rank	Source
SMT 2025	91%	#6 of 40	Official leaderboard
AIME 2025	94.5%	#24 of 143	Model card
CMIMC 2025	90.6%	#6 of 33	Official leaderboard
GPQA Diamond	84.2%	#74 of 219	Epoch AI
BRUMO 2025	93.3%	#15 of 42	Official leaderboard
FrontierMath	21.4%	#28 of 70	Epoch AI
OTIS Mock AIME 2024-2025	83.1%	#84 of 202	Epoch AI
HiPhO	57.2%	#16 of 35	Official leaderboard
HMMT 2025	89.4%	#23 of 49	Model card
General365	53%	#13 of 23	Official leaderboard
HMMT Nov 2025	89.2%	#14 of 22	Official leaderboard
SimpleBench	39.6%	#70 of 100	Official leaderboard
IMO-AnswerBench	78.6%	#27 of 37	Model card
FrontierMath Tier 4	0%	#52 of 55	Epoch AI
MathArena (Apex, IMO, USAMO)	0%	#47 of 47	Official leaderboard

Benchmark	Score	Rank	Source
LiveCodeBench	83.1%	#8 of 73	Model card
METR Time Horizon	58 min	#10 of 31	Official leaderboard
OJBench	48.7%	#4 of 12	Model card
SciCode	44.8%	#19 of 47	Model card
SWE-bench Verified	63.4%	#27 of 60	Official leaderboard
Multi-SWE-bench	41.9%	#6 of 10	Model card
WeirdML	42.8%	#68 of 112	Official leaderboard
Bird-SQL	59.9%	#8 of 13	Official leaderboard
ALE-Bench	598	#94 of 121	Epoch AI
OTelBench	7.25%	#11 of 14	Official leaderboard
WebDev Arena	1,337	#85 of 105	Epoch AI
SWE-bench Multilingual	61.1%	#58 of 70	Model card
OpenHands Index	41%	#28 of 33	Official leaderboard
PostTrainBench	7.25%	#23 of 23	Official leaderboard

Benchmark	Score	Rank	Source
Terminal-Bench	47.1%	#3 of 22	Model card
SnakeBench	29	#41 of 183	Official leaderboard
Step Race	3	#15 of 59	Official leaderboard
Elimination Game	4	#26 of 56	Official leaderboard
Terminal-Bench 2.0	35.7%	#27 of 43	Official leaderboard
BrowseComp-ZH	62.3%	#11 of 17	Model card
BrowseComp	60.2%	#54 of 77	Model card
Pencil Puzzle Bench	0%	#39 of 53	Official leaderboard
BalatroBench	5.47	#12 of 16	Official leaderboard
Agent Poker Bench	1,012	#14 of 17	Official leaderboard
VITA-Bench	12.8%	#15 of 18	Official leaderboard
Toolathlon	17.6%	#42 of 50	Official leaderboard
RWSearch	63%	#7 of 7	Official leaderboard

Benchmark	Score	Rank	Source
CaseLaw v2	65.7%	#11 of 53	Official leaderboard
ProfBench	48.9%	#32 of 72	Official leaderboard
HealthBench	58%	#16 of 30	Model card
TaxEval v2	71.7%	#71 of 129	Official leaderboard
CorpFin v2	60.6%	#67 of 119	Official leaderboard
Vals Finance Agent v1	26%	#28 of 47	Official leaderboard
LegalBench	80.2%	#85 of 138	Official leaderboard
PRBench	40.9%	#30 of 44	Official leaderboard
Physical-Engineering's Last Exam	0.06	#13 of 18	Official leaderboard
Vals Finance Agent	36.6%	#41 of 49	Official leaderboard
prinzbench	22.0	#22 of 26	Official leaderboard
FinSearchComp	47.4%	#7 of 7	Model card

Benchmark	Score		Rank	Source
Human Pathogen Capabilities Test v2 (HPCT v2)	50.4%		#42 of 71	Official leaderboard
Agentic BAIM-LLM Evaluation (ABLE)	48.1%		#10 of 16	Official leaderboard

Benchmark	Score		Rank	Source
CAIS Text Capabilities Index	19.9		#49 of 66	Official leaderboard
CAIS Risk Index	63.9		#53 of 65	Official leaderboard

Moonshot AI's models in color, against every other model on the Epoch Capabilities Index, by release date.

The benchmarks the lab chose to headline, with the scores it gave.