Back to Models
DeepSeekOpen Source
DeepSeek-V3.2-Exp
DeepSeek-V3.2-Exp, developed by DeepSeek-AI, is an experimental version of a foundation model that first introduces DeepSeek Sparse Attention (DSA). While maintaining performance equivalent to V3.1-Terminus, it significantly improves inference speed in long-context scenarios.
Parameters
Undisclosed
Context Window
License
MIT
Release Date
2025-09-29
API Pricing
API pricing for this model is not yet available
Strengths
- ・Fast inference for long-context
- ・Efficient attention via DSA architecture
- ・Significantly reduced API costs
Weaknesses
- ・Experimental version status
- ・Performance matches previous version
- ・Instability as a non-final release
Use Cases
- ・Long-context document analysis
- ・Cost-optimized API implementations
- ・Dialogues requiring fast responses
Deep Analysis
Parameters
685B MoE
Architecture
DeepSeek Sparse Attention (DSA)
Pricing
$0.14/$0.21 per 1M tokens
50% cheaper than V3.1-Terminus
Release Date
September 29, 2025
Strengths
- ・Introduces DSA (O(L²)→O(Lk))
- ・50%+ price reduction vs V3.1-Terminus
- ・Improved long-context efficiency
- ・Open-source with GPU kernels
Weaknesses
- ・Experimental, not production
- ・Comparable to V3.1-Terminus
- ・Superseded by V3.2
- ・Limited benchmark data
Competitor Comparison
| Model | SWE | Price |
|---|---|---|
| V3.1-Terminus | 66.0% | $0.30/$0.95 |
| V3.2 | 73.1% | $0.28/$0.42 |
DeepSeek-V3.2-Exp introduced DeepSeek Sparse Attention (DSA), reducing attention complexity from O(L²) to O(Lk). 50%+ cheaper than V3.1-Terminus.
Sources
Analysis generated: 2026-05-24