Which is better, DeepSeek-V3.2 or Kimi K2.5?

Across 65 shared benchmarks tracked on Vector Wire, DeepSeek-V3.2 leads on 14 and Kimi K2.5 leads on 51 — Kimi K2.5 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 K2.5

DeepSeekvsMoonshot65 shared benchmarks14–51 head-to-head

Benchmark	DeepSeek-V3.2	Kimi K2.5
AA Agentic Index	39.8	52.8
AA Intelligence	32.8	36
AA-LCR	70.7	73
AA-Omniscience	-22.5	-7.3
AIME 2025	93.1	96.1
AIME 2026 I	92.7	92.5
aime_2026	95.1	95.8
apexAgents	14.5	11.5
arc_agi_1	57	65.3
ARC-AGI-2	4	11.8
arena_text_factuality	1433	1445
Artificial Analysis Coding Index	44.2	46.8
browsecomp	51.4	74.9
BrowseComp (context management)	67.6	74.9
browsecomp_with_context_manager	67.6	74.9
browsecomp_zh	65	62.3
coding_arena_elo	1368	1436
critpt	2.9	3.1
cybergym	17.3	41.3
deepsearchqa_f1	60.9	89
FinSearchCompT2&T3	59.1	67.8
frontiermath_tier_4	2.1	4.2
gaia_no_file	75.1	75.9
gdpval	18.8	38.3
GPQA Diamond	84	87.9
HLE	40.8	50.2
HLE (with tools)	40.8	51.8
HMMT 2025	90.2	95.4
hmmt_feb_2025	92.5	95.4
hmmt_feb_2026	84.1	87.1
hmmt_nov_2025	90.2	91.1
IFBench	61	70.2
imo_answer_bench	78.3	81.8
LiveCodeBench v6	83.3	85
longbench_v2	59.8	61
MCP Atlas	62.2	64
mcpmark	38	29.5
MMLU-Pro	86	87.1
OJ-Bench (cpp)	54.7	57.4
OmniScience Accuracy	33	35.2
OmniScience Non-Hallucination	17.3	50
PaperBench	47.1	63.5
ResearchRubrics	55.8	59.5
scicode	39	49
Seal-0	49.5	57.4
simpleqa_verified	27.5	36.9
swe_bench_bash	60	70.8
SWE-bench Multilingual	70.2	73
SWE-bench Pro	15.6	53.8
SWE-bench Verified	73.1	76.8
TauBench V3 - Banking	18.8	14.2
Terminal-Bench 2.0	46.4	50.8
Terminal-Bench 2.1	46.8	45.7
Terminal-Bench Hard	35.6	34.8
tool_decathlon	35.2	27.8
toolathlon	35.2	27.8
vectara_answer_rate	92.6	92.2
vectara_avg_summary_length	62	112
vectara_factual_consistency	93.7	85.8
vectara_hallucination_rate ↓	6.3	14.2
WideSearch (item-f1)	32.5	72.7
xbench_deepsearch	71	76.7
τ²-Bench	91	85.4
τ²-Bench Telecom (AA run)	90.6	95.9
τ³-Bench	69.2	66

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.