DeepSeek Goes All Out! V3.2 Released Before the Holiday, API Costs Slashed by 50%
As the double holidays approach, some people are already on their way for a break — while others are still watching DeepSeek roll out new models.
On September 29 (Beijing time), traces of DeepSeek V3.2 suddenly appeared on Hugging Face, sparking speculation among developers: “Is DeepSeek about to release a new model?” The suspense didn’t last long. Moments later, the official announcement of DeepSeek V3.2-Exp arrived — with a special note that this is an experimental version. The release brings two major highlights:
- Based on DeepSeek V3.1-Terminus, it introduces DeepSeek Sparse Attention (DSA) — a sparse attention mechanism that improves efficiency in long-text training and inference. It’s the first to implement fine-grained sparse attention, achieving a significant boost in performance for long-context processing with minimal impact on output quality.
- A huge benefit for developers: API pricing has been dramatically reduced. DeepSeek announced that the cost of calling its API is now more than 50% lower.
The official app, web version, and mini program have now all been updated to DeepSeek-V3.2-Exp, and developers can start trying it out right away.
However, since this is still an experimental version, DeepSeek emphasized that broader, real-world testing across different use cases is needed to ensure consistent performance and eliminate potential weak spots.
Despite the excitement, there’s also been a fair share of groans and curiosity in the community.
In the DeepSeek V3.1-Terminus community, one user commented: “It’s National Day holiday — please give us a break!” And right after V3.2-Exp was officially released, the first reactions were: “Did it really have to drop right before the holiday?”
As new models continue to roll out during the holidays, what’s your take on this wave of updates?
Original Source: https://blog.csdn.net/csdnnews/article/details/152307110