DeepSeek iconDeepSeek: DeepSeek V3.2 Exp

Model Type

Proprietary model icon

Proprietary Model

API access only

Recommended Use Cases

Text Generation

Try DeepSeek V3.2 Exp

An experimental version of DeepSeek-V3.2, serving as an intermediate step toward next-generation architecture (October 2025). V3.2-Exp introduces DeepSeek Sparse Attention while maintaining performance parity with V3.1-Terminus.

Per DeepSeek:

DeepSeek-V3.2-Exp builds upon V3.1-Terminus by introducing DeepSeek Sparse Attention—a sparse attention mechanism designed to explore and validate optimizations for training and inference efficiency in long-context scenarios.

Role in V3.2 Series

V3.2-Exp is a research-oriented release focused on validating the DeepSeek Sparse Attention architecture rather than advancing raw task accuracy. Training configurations were deliberately aligned with V3.1-Terminus to enable direct comparison.

Key Features

  • Architecture: 685B MoE with DeepSeek Sparse Attention (DSA)
  • Context Window: 128K tokens
  • Purpose: Architecture validation and research
  • License: MIT

DeepSeek Sparse Attention (DSA)

  • Fine-grained sparse attention mechanism
  • Substantial improvements in long-context training and inference efficiency
  • Maintains virtually identical model output quality

Links