#23
DeepSeek V3.2
deepseek-v3.2-exp

DeepSeek V3.2 Exp is an experimental model introducing the new DeepSeek Sparse Attention (DSA) mechanism for superior long-context efficiency. Built on the V3.1-Terminus architecture, it delivers high-performance inference at significantly reduced costs. This model excels in general-purpose tasks and coding, offering a massive 160k context window while maintaining competitive benchmark scores against proprietary alternatives.

Performance Metrics

Helpfulness
Instruction Following
Comprehension
Empathy
Creative Writing
83.40
Helpfulness
86.5
Empathy
79
Instruction Following
88.5
Creative Writing
76
Comprehension
87
Speed
Avg 95 tok/s

Model Specifications

Release Date
September 29, 2025
Lab
DeepSeek
Type
Open Source
Context Size
160K
Max Output Tokens
8.2K
Cost per 1M tokens
$0.07 / $0.14
Model Inputs
Text
Model Outputs
Text
Tool Calling
Enabled

Compare With Similar Models