DeepSeek-V3.2-Exp (DeepSeek Sparse Attention)
First production use of DeepSeek Sparse Attention, with V3.1-Terminus-level quality, lower long-context cost and API prices cut by over 50%.
A lightning indexer selects the top-k tokens each query attends to, replacing dense attention; training config was deliberately matched to V3.1-Terminus to isolate DSA's effect. Benchmarks tie with Terminus (MMLU-Pro 85.0 both).
- Date
- Monday, 29 September 2025
- Lab
- DeepSeek
- Kind
- open-weights
- Access
- open weights
- Price
- API prices reduced by more than 50% (2025-09-29)
Figures
| Measure | Value | Measured by |
|---|---|---|
| API price cut | >50% Quality is at parity, with MMLU-Pro 85.0 for both and AIME 2025 89.3 vs 88.4 | company |
Labeled experimental; V3.1-Terminus stayed available until 2025-10-15 for comparison. The 'first time' fine-grained sparse attention claim is DeepSeek's own.
Sources
- api-docs.deepseek.com/news/news250929
- huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
- github.com/deepseek-ai/DeepSeek-V3.2-Exp
This record was checked against its sources on 6 October 2026. How we check