DeepSeek V3.2 is a large-scale Mixture-of-Experts language model that harmonizes high computational efficiency with frontier-level reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained mechanism that reduces attention complexity from quadratic to linear, significantly cutting training and inference costs in long-context scenarios. Through scalable reinforcement learning post-training, it achieves performance comparable to GPT-5, with gold-medal results on the 2025 International Mathematical Olympiad and Olympiad in Informatics. The model also features a large-scale agentic task synthesis pipeline that improves instruction following and tool use in complex interactive environments.
API|Reasoning|Open ModelMIT
Knowledge Cutoff
2025-03
Input → Output Format
Context Memory
164KIN164KOUT
AI Performance Evaluation
Arena Overall Score
1424
±4As of 2026-05-01
Overall Rank
No.71
44,809 Votes
Arena by Ability
Hard Prompts
1447±5No.69
Expert Knowledge
1447±11No.71
Instruction Following
1419±6No.63
Conversation Memory
1427±7No.71
Creative
1399±8No.66
Coding
1468±7No.72
Math
1429±11No.63
Arena by Occupation
Creative Writing
1409±7No.64
Social Sciences
1447±8No.67
Media
1394±7No.74
Business
1420±7No.74
Healthcare
1441±12No.81
Legal
1430±11No.76
Software
1456±6No.75
Mathematics
1438±13No.62
Source:Arena Intelligence
Overall
AA Intelligence Index
42%↑3%
LiveBench
50%↓11%
Reasoning & Math
AA Math Index
92%↑18%
GPQA Diamond
84%↑2%
HLE
22%↑5%
MMLU-Pro
86%↑5%
AIME 2025
92%↑18%
LB Reasoning
44%↓25%
LB Math
64%↓10%
LB Data
45%↓8%
Coding
AA Coding Index
37%↑0%
LiveCodeBench
86%↑21%
LB Coding
76%↑3%
LB Agentic
47%↑2%
TAU2
91%↑10%
TerminalBench
36%↑2%
SciCode
39%↓3%
Language & Instructions
IFBench
61%↓2%
AA-LCR
65%↑3%
Hallucination (HHEM)
6.1%↓4%
Factual (HHEM)
94%↑4%
LB Language
64%↓8%
LB IF
23%↓28%
Output Speed
Standard Mode
47tok/s↓30
First Output 1.26s
Reasoning Mode
79tok/s↓7
First Output 25.96s