DeepSeek-V3.2 (Thinking)

DeepSeek-V3.2 in thinking mode. A powerful reasoning model with 685B parameters using DeepSeek Sparse Attention (DSA).

AIME 2025

93.1%

i
HMMT 2025

90.2%

i
MMLU-Pro

85.0%

i
LiveCodeBench

83.3%

i
GPQA

82.4%

i
t2-bench

80.2%

i
CodeForces

79.5%

i
SWE-Bench Verified

73.1%

i
SWE-bench Multilingual

70.2%

i
BrowseComp-zh

65.0%

i
BrowseComp

51.4%

i
Terminal-Bench 2.0

46.4%

i
Toolathlon

35.2%

i
Humanity's Last Exam

25.1%

i