DeepSeek Sparse Attention (V3.2-Exp)
DeepSeek Sparse Attention (DSA) in V3.2-Exp is fine-grained sparse attention with output quality near V3.1-Terminus, which enabled a 50%+ API price cut.
A trained 'lightning indexer' selects which tokens each query attends to, cutting long-context compute from quadratic toward linear while preserving quality. Released as an experimental model with open weights and an immediate API price cut.
- Date
- Monday, 29 September 2025
- Lab
- DeepSeek
- Kind
- paper
- Access
- open weights
- Price
- API price cut >50% (2025-09-29)
Figures
| Measure | Value | Measured by |
|---|---|---|
| API price cut | >50% effective immediately at release | company |
Release-notes date 2025-09-29. The 'lightning indexer' mechanism detail is from the V3.2 paper I did not fully open; the arXiv V3.2 report followed on 2025-12-02.
Sources
This record was checked against its sources on 6 October 2026. How we check