Which is better, DeepSeek-V4-Flash or Qwen3.5 397B A17B?

Across 47 shared benchmarks tracked on Vector Wire, DeepSeek-V4-Flash leads on 33 and Qwen3.5 397B A17B leads on 13 — DeepSeek-V4-Flash leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

DeepSeek-V4-Flash vs Qwen3.5 397B A17B

DeepSeekvsAlibaba47 shared benchmarks33–13 head-to-head

Benchmark	DeepSeek-V4-Flash	Qwen3.5 397B A17B
AA Agentic Index	61.3	53.3
AA Intelligence	51.8	34.3
AA-LCR	74.3	72.7
AA-Omniscience	-14.3	-30.8
aime_2026	95.8	94.2
arena_elo	1439	1442
arena_text_factuality	1434	1445
Artificial Analysis Coding Index	69.1	48.2
browsecomp	73.2	69
BrowseComp (context management)	73.2	78.6
BrowseComp (with context management)	73.2	78.6
BrowseComp with context management	73.2	78.6
coding_arena_elo	1431	1399
critpt	16.6	1.7
FORTRESS (adversarial)	32	77.3
FORTRESS (benign)	99.2	95.4
gdpval	52.9	35.8
GDPval-AA v2	1189	962
Global-MMLU-Lite	88.4	90
GPQA Diamond	90.8	89.3
HLE	38.6	29
HLE (with tools)	45.1	48.3
hmmt_feb_2026	94.8	87.9
IFBench	79.2	78.8
imo_answer_bench	88.4	80.9
itbenchSre	31.5	34.1
MMLU-Pro	86.4	87.8
nl2repo	54.2	32.2
OmniScience Accuracy	40.4	30.8
OmniScience Non-Hallucination	15.6	17.3
scicode	49.9	42
simpleqa_verified	45	26
strongreject	97.4	99.4
SWE-bench Multilingual	73.3	69.3
SWE-bench Pro	52.6	50.9
SWE-bench Verified	79	76.4
SWEBench Pro (public)	52.6	50.9
SWEBench Pro Public	52.6	50.9
Tau 3 Banking	22.9	13.4
TauBench V3 - Banking	39.4	13.4
Terminal Bench 2.1 (best harness)	61.8	51.3
Terminal-Bench 2.0	56.9	52.5
Terminal-Bench 2.1	82.7	51.3
Terminal-Bench Hard	38.6	40.9
toolathlon	70.3	38.3
τ²-Bench Telecom (AA run)	95.6	95.6
τ³-Bench	22.9	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.