Which is better, Gemma 4 31B or Qwen3.5 35B A3B?

Across 69 shared benchmarks tracked on Vector Wire, Gemma 4 31B leads on 28 and Qwen3.5 35B A3B leads on 40 — Qwen3.5 35B A3B 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 Qwen3.5 35B A3B

GooglevsAlibaba69 shared benchmarks28–40 head-to-head

Benchmark	Gemma 4 31B	Qwen3.5 35B A3B
AA Agentic Index	14.4	44.1
AA Intelligence	29.7	29.9
AA-LCR	68.3	68.3
AA-Omniscience	-47.9	-48.1
AI2D	89	92.6
aime_2026	89.2	93.3
Artificial Analysis Coding Index	43.4	37
C-Eval	82.6	90.2
CC-OCR	75.7	80.7
charxiv_rq	67.9	77.5
Claw Eval (pass@3)	25	51
Claw-Eval Avg	48.5	65.4
coding_arena_elo	1364	1250
critpt	1.4	0.9
DeepPlanning	24	22.8
ERQA	57.5	64.8
gdpval	15.5	20.3
GPQA (unspecified)	84.3	84.2
GPQA Diamond	85.7	84.5
HallusionBench	67.4	67.9
HLE	26.5	47.4
hmmt_feb_2025	88.7	89
hmmt_feb_2026	77.2	81.8
hmmt_nov_2025	87.5	89.2
IFBench	75.6	72.5
imo_answer_bench	74.5	76.8
itbenchSre	37.3	21.5
LiveCodeBench v6	80	74.6
mathvision	85.6	83.9
MathVista	79.3	86.2
MCP Atlas	57.2	62.4
mcpmark	18.1	27
MMBench EN-DEV-v1.1	90.9	91.5
mmlu_redux	93.7	93.3
MMLU-Pro	85.2	85.3
mmmlu	88.4	85.2
MMMU	80.4	81.4
MMMU-Pro	76.9	75.1
MMStar	77.3	81.9
nl2repo	15.5	20.5
OCRBench	86.1	91
OmniDocBench 1.5	80.1	89.3
OmniScience Accuracy	20	20.1
OmniScience Non-Hallucination	18.1	14.6
QwenClawBench	41.7	47.7
QwenWebBench	1197	978
RealWorldQA	72.3	84.1
RefSpatial-Bench	4.7	0.6
scicode	43.4	37.7
SimpleVQA	52.9	58.3
SkillsBench Avg5	23.6	4.4
supergpqa	65.7	63.4
SWE-bench Multilingual	51.7	60.3
SWE-bench Pro	35.7	44.6
SWE-bench Verified	52	70
TauBench V3 - Banking	14.8	4.9
Terminal-Bench 2.0	42.9	40.5
Terminal-Bench 2.1	43.4	40.8
Terminal-Bench Hard	36.4	26.5
tool_decathlon	21.2	28.7
vectara_answer_rate	100	99.8
vectara_avg_summary_length	75.8	94.9
vectara_factual_consistency	92.6	89.5
vectara_hallucination_rate ↓	7.4	10.5
VITA-Bench	43	31.9
WideSearch	35.2	59.1
ZEROBench_sub	26	34.1
τ²-Bench Telecom (AA run)	65.5	89.2
τ³-Bench	67.5	68.9

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.