Which is better, Qwen3.5 397B A17B or Qwen3.6 Plus?

Across 65 shared benchmarks tracked on Vector Wire, Qwen3.5 397B A17B leads on 24 and Qwen3.6 Plus leads on 39 — Qwen3.6 Plus 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 397B A17B vs Qwen3.6 Plus

AlibabavsAlibaba65 shared benchmarks24–39 head-to-head

Benchmark	Qwen3.5 397B A17B	Qwen3.6 Plus
AA Agentic Index	53.3	29
AA Intelligence	34.3	40.5
AA-LCR	72.7	72.3
AA-Omniscience	-30.8	2.6
AgentWorldBench - Android	54.9	57.6
AgentWorldBench - MCP	68.3	55.3
AgentWorldBench - OS	60.9	60.3
AgentWorldBench - Overall	54.7	50.8
AgentWorldBench - Search	30.8	21.9
AgentWorldBench - SWE	64.4	59.1
AgentWorldBench - Terminal	55.3	50.6
AgentWorldBench - Web	48.5	50.8
aime_2026	94.2	95.3
arena_elo	1442	1444
arena_text_factuality	1445	1438
Artificial Analysis Coding Index	48.2	54.5
C-Eval	93	93.3
Claw Eval (pass@3)	48.1	58.7
coding_arena_elo	1399	1460
critpt	1.7	2.9
DeepPlanning	34.3	41.5
ERQA	67.5	65.7
gdpval	35.8	32
GPQA Diamond	89.3	90.4
HLE	29	28.8
HLE (with tools)	48.3	50.6
HMMT 2025	92.7	94.6
hmmt_feb_2026	87.9	87.8
hmmt_nov_2025	92.7	94.6
IFBench	78.8	75.2
ifeval	92.6	94.3
imo_answer_bench	80.9	83.8
include	85.6	85.1
LiveCodeBench v6	83.6	87.1
longbench_v2	63.2	62
mcpmark	46.1	48.2
MLVU	86.7	86.7
mmlu_prox	84.7	84.7
mmlu_redux	94.9	94.5
MMLU-Pro	87.8	88.5
mmmlu	88.5	89.5
MMMU	85	86
MMMU-Pro	79	78.8
MMStar	83.8	83.3
nl2repo	32.2	37.9
OmniScience Accuracy	30.8	26.4
OmniScience Non-Hallucination	17.3	68
RealWorldQA	83.9	85.4
scicode	42	40.7
simpleqa_verified	26	49.1
SimpleVQA	67.1	0.7
supergpqa	70.4	71.6
SWE-bench Multilingual	69.3	73.8
SWE-bench Pro	50.9	56.6
SWE-bench Verified	76.4	78.8
TauBench V3 - Banking	13.4	20.8
Terminal-Bench 2.0	52.5	61.6
Terminal-Bench 2.1	51.3	61.4
Terminal-Bench Hard	40.9	43.9
toolathlon	38.3	39.8
VideoMMMU	84.7	84
VITA-Bench	49.7	44.3
WideSearch	74	74.3
τ²-Bench Telecom (AA run)	95.6	97.7
τ³-Bench	13.4	70.7

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.