Which is better, Gemma 4 31B or Kimi K2.5?

Across 52 shared benchmarks tracked on Vector Wire, Gemma 4 31B leads on 11 and Kimi K2.5 leads on 41 — 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 —

Gemma 4 31B vs Kimi K2.5

GooglevsMoonshot52 shared benchmarks11–41 head-to-head

Benchmark	Gemma 4 31B	Kimi K2.5
AA Agentic Index	14.4	52.8
AA Intelligence	29.7	36
AA-LCR	68.3	73
AA-Omniscience	-47.9	-7.3
aime_2026	89.2	95.8
arena_elo	1451	1450
arena_text_factuality	1449	1445
arena_vision	1252	1247
Artificial Analysis Coding Index	43.4	46.8
charxiv_rq	67.9	78.7
coding_arena_elo	1364	1436
critpt	1.4	3.1
gdpval	15.5	38.3
GPQA Diamond	85.7	87.9
HLE	26.5	50.2
HLE (with tools)	26.5	51.8
hmmt_feb_2025	88.7	95.4
hmmt_feb_2026	77.2	87.1
hmmt_nov_2025	87.5	91.1
IFBench	75.6	70.2
imo_answer_bench	74.5	81.8
LiveCodeBench v6	80	85
mathvision	85.6	84.2
MathVista	79.3	90.1
MCP Atlas	57.2	64
mcpmark	18.1	29.5
MMLU-Pro	85.2	87.1
MMMU-Pro	76.9	78.5
nl2repo	15.5	32
OCRBench	86.1	92.3
OmniDocBench 1.5	80.1	88.8
OmniScience Accuracy	20	35.2
OmniScience Non-Hallucination	18.1	50
scicode	43.4	49
simpleqa_verified	9.6	36.9
SimpleVQA	52.9	71.2
SWE-bench Multilingual	51.7	73
SWE-bench Pro	35.7	53.8
SWE-bench Verified	52	76.8
TauBench V3 - Banking	14.8	14.2
Terminal-Bench 2.0	42.9	50.8
Terminal-Bench 2.1	43.4	45.7
Terminal-Bench Hard	36.4	34.8
tool_decathlon	21.2	27.8
vectara_answer_rate	100	92.2
vectara_avg_summary_length	75.8	112
vectara_factual_consistency	92.6	85.8
vectara_hallucination_rate ↓	7.4	14.2
WideSearch	35.2	79
τ²-Bench	76.9	85.4
τ²-Bench Telecom (AA run)	65.5	95.9
τ³-Bench	67.5	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.