Which is better, DeepSeek-V3.2 or Gemma 4 31B?

Across 43 shared benchmarks tracked on Vector Wire, DeepSeek-V3.2 leads on 32 and Gemma 4 31B leads on 10 — DeepSeek-V3.2 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 Gemma 4 31B

DeepSeekvsGoogle43 shared benchmarks32–10 head-to-head

Benchmark	DeepSeek-V3.2	Gemma 4 31B
AA Agentic Index	39.8	14.4
AA Intelligence	32.8	29.7
AA-LCR	70.7	68.3
AA-Omniscience	-22.5	-47.9
aime_2026	95.1	89.2
arena_text_factuality	1433	1449
Artificial Analysis Coding Index	44.2	43.4
coding_arena_elo	1368	1364
critpt	2.9	1.4
gdpval	18.8	15.5
GPQA (unspecified)	79.9	84.3
GPQA Diamond	84	85.7
HLE	40.8	26.5
HLE (with tools)	40.8	26.5
hmmt_feb_2025	92.5	88.7
hmmt_feb_2026	84.1	77.2
hmmt_nov_2025	90.2	87.5
IFBench	61	75.6
imo_answer_bench	78.3	74.5
LiveCodeBench v6	83.3	80
MCP Atlas	62.2	57.2
mcpmark	38	18.1
mmlu_redux	93.7	93.7
MMLU-Pro	86	85.2
OmniScience Accuracy	33	20
OmniScience Non-Hallucination	17.3	18.1
scicode	39	43.4
simpleqa_verified	27.5	9.6
SWE-bench Multilingual	70.2	51.7
SWE-bench Pro	15.6	35.7
SWE-bench Verified	73.1	52
TauBench V3 - Banking	18.8	14.8
Terminal-Bench 2.0	46.4	42.9
Terminal-Bench 2.1	46.8	43.4
Terminal-Bench Hard	35.6	36.4
tool_decathlon	35.2	21.2
vectara_answer_rate	92.6	100
vectara_avg_summary_length	62	75.8
vectara_factual_consistency	93.7	92.6
vectara_hallucination_rate ↓	6.3	7.4
τ²-Bench	91	76.9
τ²-Bench Telecom (AA run)	90.6	65.5
τ³-Bench	69.2	67.5

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.