Which is better, DeepSeek-V3 or Qwen2.5 Instruct 72B?

Across 94 shared benchmarks tracked on Vector Wire, DeepSeek-V3 leads on 79 and Qwen2.5 Instruct 72B leads on 14 — DeepSeek-V3 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

DeepSeek-V3 vs Qwen2.5 Instruct 72B

DeepSeekvsAlibaba94 shared benchmarks79–14 head-to-head

Benchmark	DeepSeek-V3	Qwen2.5 Instruct 72B
AA Intelligence	15.4	10
AA-LCR	41.3	20.3
AA-Omniscience	-37.6	-52.2
AGIEval	79.6	75.8
aider_edit_acc	79.7	65.4
aider_polyglot	55.1	7.6
AIME 2024	59.4	23.3
AIR-Bench 2024	40.8	59
anthropic_red_team	97.1	99.6
arc_challenge	95.3	94.5
ARC-Challenge (Acc.)	95.3	94.5
ARC-Easy (Acc.)	98.9	98.4
Arena Hard	91.4	81.2
Artificial Analysis Coding Index	23	11.9
bbh	87.5	79.8
BBH (EM)	87.5	79.8
bbq	96.7	95.4
C-Eval	90.1	89.2
C3	78.6	76.7
C3 (Acc.)	78.6	76.7
CCPM	92	88.5
Chinese SimpleQA (C-SimpleQA)	68	52.2
CLUEWSC	90.9	91.4
CMath (EM)	90.7	84.5
cmmlu	88.8	89.5
CMMLU (Acc.)	88.8	89.5
CMRC	76.3	75.8
Codeforces	1134	24.8
critpt	0	0
CRUXEval-I (Acc.)	67.3	59.1
CRUXEval-I (input prediction)	67.3	59.1
CRUXEval-O (Acc.)	69.8	59.9
CRUXEval-O (output prediction)	69.8	59.9
DROP	91.6	76.7
DROP (3-shot F1)	91.6	76.7
DROP (F1)	91	85
FRAMES (Acc.)	73.3	69.8
GPQA Diamond	68.4	49.1
GSM8K	96.7	95.8
GSM8K (EM)	89.3	88.3
harmbench	49.7	72.8
hellaswag	88.9	84.8
HellaSwag (Acc.)	88.9	84.8
HLE	5.2	4.2
humaneval	92.1	86.6
HumanEval-Mul (Pass@1)	82.6	77.3
IFBench	41	36.9
ifeval	87.3	87.2
IFEval (avg)	87.3	87.2
livebench	72.4	52.3
LiveCodeBench	49.6	55.5
livecodebench_pass1cot	40.5	31.1
LiveCodeBench-Base (Pass@1)	19.4	12.9
LongBench v2 overall (w/o CoT)	48.7	42.1
longbench_v2	48.7	39.4
MATH	91.2	88.4
MATH (EM)	61.6	54.4
math_500_em	94	80
MBPP	75.4	72.6
MBPP+ (EvalPlus-augmented)	78.8	77
mgsm	79.8	87.3
MGSM (EM)	79.8	76.2
mmlu	89.4	86.1
MMLU (Acc.)	87.1	85
mmlu_redux	90.5	86.8
MMLU-Pro	81.2	71.6
MMLU-Pro (Acc.)	64.4	58.3
MMLU-Redux (Acc.)	86.2	83.2
mmmlu	79.4	74.8
MMMLU-non-English (Acc.)	79.4	74.8
MultiPL-E	83.1	75.1
narrativeqa	79.6	74.5
NaturalQuestions	40	33.2
NaturalQuestions (EM)	40	33.2
naturalquestions_closedbook	46.7	35.9
OmniScience Accuracy	25.4	17.6
OmniScience Non-Hallucination	23.3	15.3
OpenBookQA	95.4	96.2
Pile-test (BPB)	0.5	0.6
PIQA	84.7	82.6
PIQA (Acc.)	84.7	82.6
RACE-High	51.3	50.3
RACE-Middle	67.1	68.1
scicode	35.8	26.7
simple_safety_tests	95.3	100
simpleqa	27.7	10.3
SWE-bench Verified	42	23.8
Terminal-Bench Hard	15.2	4.5
The Pile (Test, BPB)	0.5	0.6
TriviaQA (EM)	82.9	71.9
winogrande	84.9	82.3
WinoGrande (Acc.)	84.9	82.3
xstest	97.1	97.9
τ²-Bench Telecom (AA run)	47.1	34.5

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.