DeepSeek releases DeepSeek-V3.2-Exp with Sparse Attention for long-context efficiency

DeepSeek-AI has released DeepSeek-V3.2-Exp, an experimental model that introduces DeepSeek Sparse Attention (DSA) to improve training and inference efficiency in long-context scenarios.

• The model builds upon V3.1-Terminus by implementing fine-grained sparse attention for the first time.
• It maintains performance on par with V3.1-Terminus across public benchmarks while optimizing computational efficiency.
• An update on 2025.11.17 resolved a Rotary Position Embedding discrepancy in the indexer module that could degrade performance.
• The model is available under an MIT License with day-0 support for vLLM and Docker images for SGLang.

This release allows researchers to validate optimizations for transformer architectures focused on extended text sequences.