DeepSeek-V3.2-Exp is an experimental iteration introducing DeepSeek Sparse Attention (DSA) to improve long-context training and inference efficiency while keeping output quality on par with V3.1. It explores fine-grained sparse attention for extended sequence processing.
发布日期2025年9月29日
参数规模685B
上下文长度—
许可证MIT
知识截止—
Benchmarks
评测成绩
| 评测基准 | 类别 | 分数 | 来源 |
|---|---|---|---|
| SimpleQA | general reasoning | 97.1 | 来源 |
| AIME 2025 | math reasoning | 89.3 | 来源 |
| MMLU-Pro | language reasoning math general | 85.0 | 来源 |
| HMMT 2025 | math | 83.6 | 来源 |
| GPQA | reasoning general | 79.9 | 来源 |
| Aider-Polyglot | general code | 74.5 | 来源 |
| LiveCodeBench | reasoning general code | 74.1 | 来源 |
| Codeforces | math reasoning | 70.7 | 来源 |
| SWE-Bench Verified | reasoning frontend_development code | 67.8 | 来源 |
| SWE-Bench Multilingual | reasoning code | 57.9 | 来源 |
| BrowseComp-zh | reasoning search | 47.9 | 来源 |
| BrowseComp | reasoning search | 40.1 | 来源 |
| Terminal-Bench | reasoning code | 37.7 | 来源 |
| Humanity's Last Exam | general | 19.8 | 来源 |
Pricing
API 价格对比
暂无 API 价格。