Which is better, Gemma 4 26B A4B or Qwen3.5 397B A17B?

Across 59 shared benchmarks tracked on Vector Wire, Gemma 4 26B A4B leads on 3 and Qwen3.5 397B A17B leads on 56 — Qwen3.5 397B A17B 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 26B A4B vs Qwen3.5 397B A17B

GooglevsAlibaba59 shared benchmarks3–56 head-to-head

Benchmark	Gemma 4 26B A4B	Qwen3.5 397B A17B
AA Agentic Index	32.1	53.3
AA Intelligence	26.1	34.3
AA-LCR	61.7	72.7
AA-Omniscience	-50.8	-30.8
aime_2026	88.3	94.2
arena_elo	1438	1442
arena_text_factuality	1445	1445
arena_vision	1238	1249
Artificial Analysis Coding Index	39.3	48.2
browsecomp	26.3	69
C-Eval	82.5	93
CC-OCR	74.5	82
charxiv_rq	69	80.8
Claw Eval (pass@3)	28	48.1
Claw-Eval Avg	58.8	70.7
coding_arena_elo	1361	1399
critpt	0	1.7
DeepPlanning	16.2	34.3
gdpval	25.7	35.8
GDPval-AA v2	807	962
GPQA Diamond	82.3	89.3
GSM8K	77	66.7
HLE	19.3	29
HLE (with tools)	17.2	48.3
hmmt_feb_2025	91.7	94.8
hmmt_feb_2026	79	87.9
hmmt_nov_2025	87.5	92.7
IFBench	72.4	78.8
ifeval	96.4	92.6
imo_answer_bench	74.3	80.9
itbenchSre	23.6	34.1
LiveCodeBench v6	77.1	83.6
mcpmark	14.2	46.1
mmlu_redux	92.7	94.9
MMLU-Pro	85.2	87.8
mmmlu	86.3	88.5
MMMU	78.4	85
MMMU-Pro	73.8	79
nl2repo	11.6	32.2
OmniScience Accuracy	19.1	30.8
OmniScience Non-Hallucination	13.6	17.3
QwenClawBench	38.7	51.8
QwenWebBench	1178	1186
RealWorldQA	72.2	83.9
scicode	40.3	42
SimpleVQA	52.2	67.1
SkillsBench Avg5	12.3	30
supergpqa	61.4	70.4
SWE-bench Multilingual	43.4	69.3
SWE-bench Pro	13.8	50.9
SWE-bench Verified	57.4	76.4
TauBench V3 - Banking	12	13.4
Terminal-Bench 2.0	34.2	52.5
Terminal-Bench 2.1	39	51.3
Terminal-Bench Hard	25	40.9
VITA-Bench	36.9	49.7
WideSearch	38.3	74
τ²-Bench Telecom (AA run)	43.6	95.6
τ³-Bench	59	13.4

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.