Which is better, Gemma 4 31B or Qwen3.5 27B?

Across 69 shared benchmarks tracked on Vector Wire, Gemma 4 31B leads on 18 and Qwen3.5 27B leads on 51 — Qwen3.5 27B 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 27B

GooglevsAlibaba69 shared benchmarks18–51 head-to-head

Benchmark	Gemma 4 31B	Qwen3.5 27B
AA Agentic Index	14.4	54.6
AA Intelligence	29.7	34.6
AA-LCR	68.3	72.3
AA-Omniscience	-47.9	-44
AI2D	89	92.9
aime_2026	89.2	92.6
arena_vision	1252	1221
Artificial Analysis Coding Index	43.4	34.9
C-Eval	82.6	90.5
CC-OCR	75.7	81
charxiv_rq	67.9	79.5
Claw Eval (pass@3)	25	46.2
Claw-Eval Avg	48.5	64.3
coding_arena_elo	1364	1357
critpt	1.4	0.9
DeepPlanning	24	22.6
DynaMath	79.5	87.7
ERQA	57.5	60.5
gdpval	15.5	33
GPQA Diamond	85.7	85.8
HallusionBench	67.4	70
HLE	26.5	48.5
hmmt_feb_2025	88.7	92
hmmt_feb_2026	77.2	84.3
hmmt_nov_2025	87.5	89.8
IFBench	75.6	76.5
imo_answer_bench	74.5	79.9
itbenchSre	37.3	35.5
LiveCodeBench v6	80	80.7
mathvision	85.6	86
MathVista	79.3	87.8
MCP Atlas	57.2	68.4
mcpmark	18.1	36.3
MMBench EN-DEV-v1.1	90.9	92.6
mmlu_redux	93.7	93.2
MMLU-Pro	85.2	86.1
mmmlu	88.4	85.9
MMMU	80.4	82.3
MMMU-Pro	76.9	75
MMStar	77.3	81
nl2repo	15.5	27.3
OCRBench	86.1	89.4
OmniDocBench 1.5	80.1	88.9
OmniScience Accuracy	20	20.7
OmniScience Non-Hallucination	18.1	24.9
QwenClawBench	41.7	52.2
QwenWebBench	1197	1068
RealWorldQA	72.3	83.7
RefSpatial-Bench	4.7	67.7
scicode	43.4	39.5
SimpleVQA	52.9	56
SkillsBench Avg5	23.6	27.2
supergpqa	65.7	65.6
SWE-bench Multilingual	51.7	69.3
SWE-bench Pro	35.7	51.2
SWE-bench Verified	52	75
Terminal-Bench 2.0	42.9	41.6
Terminal-Bench Hard	36.4	32.6
tool_decathlon	21.2	31.5
vectara_answer_rate	100	99.8
vectara_avg_summary_length	75.8	94.4
vectara_factual_consistency	92.6	87.9
vectara_hallucination_rate ↓	7.4	12.1
VITA-Bench	43	41.9
VlmsAreBlind	87.2	96.9
WideSearch	35.2	66.4
ZEROBench_sub	26	36.2
τ²-Bench Telecom (AA run)	65.5	93.9
τ³-Bench	67.5	68.4

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.