Which is better, Qwen3.5 27B or Qwen3.5 397B A17B?

Across 68 shared benchmarks tracked on Vector Wire, Qwen3.5 27B leads on 11 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 —

Qwen3.5 27B vs Qwen3.5 397B A17B

AlibabavsAlibaba68 shared benchmarks11–56 head-to-head

Benchmark	Qwen3.5 27B	Qwen3.5 397B A17B
AA Agentic Index	54.6	53.3
AA Intelligence	34.6	34.3
AA-LCR	72.3	72.7
AA-Omniscience	-44	-30.8
aime_2026	92.6	94.2
arena_vision	1221	1249
Artificial Analysis Coding Index	34.9	48.2
bfcl_v4	68.5	72.9
browsecomp	61	69
browsecomp_zh	62.1	70.3
C-Eval	90.5	93
CC-OCR	81	82
charxiv_rq	79.5	80.8
Claw Eval (pass@3)	46.2	48.1
Claw-Eval Avg	64.3	70.7
coding_arena_elo	1357	1399
critpt	0.9	1.7
DeepPlanning	22.6	34.3
DynaMath	87.7	86.3
ERQA	60.5	67.5
gdpval	33	35.8
GPQA Diamond	85.8	89.3
HLE	48.5	29
HMMT 2025	89.8	92.7
hmmt_feb_2025	92	94.8
hmmt_feb_2026	84.3	87.9
hmmt_nov_2025	89.8	92.7
IFBench	76.5	78.8
ifeval	95	92.6
imo_answer_bench	79.9	80.9
include	81.6	85.6
itbenchSre	35.5	34.1
LiveCodeBench v6	80.7	83.6
longbench_v2	60.6	63.2
mcpmark	36.3	46.1
MLVU	85.9	86.7
mmlu_prox	82.2	84.7
mmlu_redux	93.2	94.9
MMLU-Pro	86.1	87.8
mmmlu	85.9	88.5
MMMU	82.3	85
MMMU-Pro	75	79
MMStar	81	83.8
multichallenge	60.8	67.6
MV-Bench	74.6	77.6
nl2repo	27.3	32.2
OmniScience Accuracy	20.7	30.8
OmniScience Non-Hallucination	24.9	17.3
QwenClawBench	52.2	51.8
QwenWebBench	1068	1186
RealWorldQA	83.7	83.9
scicode	39.5	42
Seal-0	47.2	46.9
SimpleVQA	56	67.1
SkillsBench Avg5	27.2	30
supergpqa	65.6	70.4
SWE-bench Multilingual	69.3	69.3
SWE-bench Pro	51.2	50.9
SWE-bench Verified	75	76.4
Terminal-Bench 2.0	41.6	52.5
Terminal-Bench Hard	32.6	40.9
V*	93.7	95.8
VideoMME (w sub.)	87	87.5
VideoMMMU	82.3	84.7
VITA-Bench	41.9	49.7
WideSearch	66.4	74
τ²-Bench Telecom (AA run)	93.9	95.6
τ³-Bench	68.4	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.