0

DeepSeek V3.2

Open weights

DeepSeek V3.2 is an AI model from DeepSeek, released with open weights.

Open
Publisher
DeepSeek
Family
Deepseek
Context
131K
Input
Output
License
mit
Released
1 Dec 2025
Canonical
huggingface.co/deepseek-ai/DeepSeek-V3.2

Cite

Notes

Only stored in your browser.

Attribution

Benchmark scores
AiderfrontiermathAALiveBenchSWE-benchmatharenaprime-hubAfterQueryVals AI
Attribution policy →
Context
131K
Input
$0.50
/1M tok
Output
$1.60
/1M tok
Blended
$0.78
#286/ 414/1M · 3:1
Speed
0
#221/ 414tok/s
Cache read
$0.03
0.05x input

Price & speed via Artificial Analysis · ranked across tracked models · caching via OpenRouter

Reported on 29 evals across 10 domains - best 85.0% on LiveBench - Math · #15 of 52

Reported eval scores

29
Agent Diff Bench74.7
Code Qa0
Infraresolutionbench68
MMLU-Pro83.7
Humanity's Last Exam (HLE)10.5
LiveBench65.9
Aider Polyglot Benchmark74.2
LiveCodeBench59.3
SWE-bench Verified70
Terminal-Bench (Hard)32.6
LiveBench - Language70.4
LiveBench - Math85
IFBench49
LiveBench - Reasoning77.2
LiveBench - Coding64.6
LiveBench - Data Analysis50
LiveBench - Instruction Following48.2
SWE-bench Multilingual59
SciCode38.7
GPQA Diamond75.1
τ²-bench (Tau²-bench)78.9
AIME 2025: Problems from the American Invitational Mathematics Examination59
ProofBench4
FrontierMath22.1
MathArena57.5
IDE-Bench31.3
Vibe Code Bench v1.15.1
TaxEval v268.2
CorpFin v247.9