Which is better, Kimi K2 Instruct or Kimi K2 Thinking?

Refresh Models —Deals —Regulatory —Sources —

Kimi K2 Instruct vs Kimi K2 Thinking

53 SHARED BENCHMARKS

Across 53 shared benchmarks, Kimi K2 Instruct scores higher on 2 and Kimi K2 Thinking on 51. The widest gap is BrowseComp w/ tools, where Kimi K2 Thinking scores 60.2 against 7.4.

MOONSHOTVSMOONSHOT53 SHARED2–51 HEAD-TO-HEAD

BenchmarkKimi K2 InstructMARGINKimi K2 Thinking

AA Agentic Index37.747.9

AA Intelligence2433.5

AA-LCR53.770.3

AA-Omniscience-26.6-21.4

aider_polyglot6059.1

AIME 20255794.5

AIME25 no tools5194.5

AIME25 w/ python75.299.1

Artificial Analysis Coding Index25.934.8

browsecomp14.160.2

BrowseComp w/ tools7.460.2

browsecomp_zh28.862.3

BrowseComp-ZH w/ tools22.262.3

critpt02.6

FinSearchComp-T310.447.4

FinSearchComp-T3 w/ tools10.447.4

Frames58.187

Frames w/ tools58.187

gdpval18.224.5

GPQA Diamond7784.5

HealthBench43.858

HealthBench no tools43.858

HLE7.923.9

HLE (with tools)26.944.9

HMMT 202538.889.4

HMMT25 no tools38.889.4

HMMT25 w/ python70.495.1

IFBench4268.1

imo_answer_bench45.878.6

LiveCodeBench6179.2

LiveCodeBench v656.183.1

LiveCodeBenchV6 no tools56.183.1

Longform Writing eval (Kimi K2 Thinking system card)62.873.8

mmlu_redux92.794.4

MMLU-Pro82.584.6

Multi-SWE-Bench33.541.9

multichallenge54.166.4

OJ-Bench (cpp)25.548.7

OJ-Bench (cpp) no tools25.548.7

OmniScience Accuracy25.430.9

OmniScience Non-Hallucination30.324.4

scicode3144.8

Seal-025.256.3

Seal-0 w/ tools25.256.3

swe_bench_bash43.863.4

SWE-bench Multilingual55.961.1

SWE-bench Verified69.271.3

Terminal-Bench44.547.1

Terminal-Bench Hard23.531.1

Terminal-Bench w/ simulated tools (JSON)44.547.1

xbench_deepsearch6176

τ²-Bench7374.3

τ²-Bench Telecom (AA run)73.493

Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.