DeepSeek-v3.2-Exp
6 hours ago
- #DeepSeek
- #AI Research
- #Sparse Attention
- DeepSeek-V3.2-Exp is an experimental version introducing DeepSeek Sparse Attention (DSA) for improved efficiency in long-context scenarios.
- DSA achieves fine-grained sparse attention, enhancing training and inference efficiency while maintaining output quality.
- Performance benchmarks show DeepSeek-V3.2-Exp is on par with V3.1-Terminus across various domains.
- Updated inference demo code and Docker images are provided for quick setup and exploration.
- The model and repository are licensed under the MIT License.