Which is better, DeepSeek-V3.2 or Kimi K2 (Reasoning)?

Across 71 shared benchmarks tracked on Vector Wire, DeepSeek-V3.2 leads on 29 and Kimi K2 (Reasoning) leads on 42 — Kimi K2 (Reasoning) leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

DeepSeek-V3.2 vs Kimi K2 (Reasoning)

DeepSeekvsMoonshot71 shared benchmarks29–42 head-to-head

Benchmark	DeepSeek-V3.2	Kimi K2 (Reasoning)
AA Agentic Index	39.8	47.9
AA Intelligence	32.8	33.5
AA-LCR	70.7	70.3
AA-Omniscience	-22.5	-21.4
AIME 2025	93.1	94.5
AIME25 no tools	89.3	94.5
AIME25 w/ python	58.1	99.1
Arena-Hard (Creative Writing)	88.8	80.1
Arena-Hard (Hard Prompt)	53.4	71.9
ArtifactsBench	55.8	54.2
Artificial Analysis Coding Index	44.2	34.8
browsecomp	51.4	60.2
BrowseComp w/ tools	40.1	60.2
browsecomp_with_context_manager	67.6	60.2
browsecomp_zh	65	62.3
BrowseComp-ZH w/ tools	47.9	62.3
critpt	2.9	2.6
FinSearchComp-global	26.2	29.5
FinSearchComp-T3	27	47.4
FinSearchComp-T3 w/ tools	27	47.4
Frames	80.2	87
Frames w/ tools	80.2	87
frontiermath_tier_4	2.1	0
GAIA (text only)	63.5	60.2
gaia_no_file	75.1	75.6
gdpval	18.8	24.5
GPQA (unspecified)	79.9	74.2
GPQA Diamond	84	84.5
HealthBench	46.9	58
HealthBench no tools	46.9	58
HLE	40.8	23.9
HLE (with tools)	40.8	44.9
HMMT 2025	90.2	89.4
hmmt_feb_2025	92.5	93.3
hmmt_nov_2025	90.2	89.2
HMMT25 no tools	83.6	89.4
HMMT25 w/ python	49.5	95.1
IFBench	61	68.1
imo_answer_bench	78.3	78.6
LiveCodeBench	86	79.2
LiveCodeBench v6	83.3	83.1
LiveCodeBenchV6 no tools	74.1	83.1
longbench_v2	59.8	45.1
Longform Writing eval (Kimi K2 Thinking system card)	72.5	73.8
mmlu_redux	93.7	94.4
MMLU-Pro	86	84.6
mrcr	55.5	44.2
Multi-SWE-Bench	37.4	41.9
OJ-Bench (cpp)	54.7	48.7
OJ-Bench (cpp) no tools	38.2	48.7
OmniScience Accuracy	33	30.9
OmniScience Non-Hallucination	17.3	25.8
ResearchRubrics	55.8	56.2
scicode	39	44.8
Seal-0	49.5	56.3
Seal-0 w/ tools	38.5	56.3
swe_bench_bash	60	63.4
SWE-bench Multilingual	70.2	61.1
SWE-bench Verified	73.1	71.3
Terminal-Bench	37.7	47.1
Terminal-Bench 2.0	46.4	35.7
Terminal-Bench Hard	35.6	31.1
Terminal-Bench w/ simulated tools (JSON)	37.7	47.1
theagentcompany	34	30
vectara_answer_rate	92.6	98.6
vectara_avg_summary_length	62	59.2
vectara_factual_consistency	93.7	82.1
vectara_hallucination_rate ↓	6.3	17.9
xbench_deepsearch	71	76
τ²-Bench	91	74.3
τ²-Bench Telecom (AA run)	90.6	93

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.