DeepSeek V3.2

Hybrid-reasoning DeepSeek model with thinking and non-thinking modes, sparse attention, and tool-use

ECI STANDING#68 of 266
SCORE146.19144.45–147.44 CI
FRONTIER LAGLaunched 7.5 months behindScores about what o3 did in April 2025.
SCORES LIKE THESE MODELS
MiniMax-M3MiniMax · Jun 2026146.47
Grok 4xAI · Jul 2025146.45
Nemotron 3 UltraNVIDIA · Jun 2026146.27
DeepSeek-V4-FlashDeepSeek · Apr 2026146.11
Full Epoch Capabilities Index ranking

From the lab

Launch announcement · Dec 1, 2025DeepSeek-V3.2 Release
RELEASE DATEDec 1, 2025
WEIGHTSOpen weights
CONTEXT WINDOW128K
MAX OUTPUT64K
REASONINGYes
TOOL CALLINGYes
INPUTStext
OUTPUTStext

Specifications from Models.dev. Availability and limits may differ by provider. Benchmark scores come from Epoch AI under CC BY 4.0.