AI BriefingDeepSeek-V3.2-Exp Experimental Model Released
Key point
DeepSeek-V3.2-Exp adds DSA and cuts API prices by more than 50%.
1 / 2
Details
DeepSeek-V3.2-Exp has been released. Built on V3.1-Terminus, it introduces DeepSeek Sparse Attention (DSA) to enable faster and more efficient training and inference on long context.
It applies across App, Web, and API, and API prices have been cut by more than 50%. For comparison testing, V3.1-Terminus remains available via a temporary API until October 15, 2025, 15:59 UTC.
Key details are as follows.
- DSA uses fine-grained sparse attention to improve long-context performance and compute cost while minimizing impact on output quality.
- On benchmarks, V3.2-Exp performs at a level similar to V3.1-Terminus.
- The model and technical report have been open-sourced, along with core GPU kernels for TileLang and CUDA.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
DeepSeek Updates Peak/Off-Peak Pricing Policy