September 29th.DeepSeek Officially released today DeepSeek-V3.2-Exp modelIt's an experimental version。
As an intermediate step towards a new generation of structures, the V3.2-Exp introduced DeepSeek Sparse Attention (Note: A Rare Focusing Mechanism) on the basis of V3.1-Terminus, to explore optimization and validation of the training and reasoning efficiency of long texts。
DeepSeek Sparse Attention (DSA) achieved, for the first time, a thin focus mechanism for fine grains, and a significant increase in long text training and reasoning efficiency, with little impact on model output。
In order to critically assess the impact of the introduction of scarce attention, the training set for DeepSeek-V3.2-Exp has been closely aligned with V3.1-Terminus. The performance of DeepSeek-V3.2-Exp in public evaluations in various fields is essentially the same as V3.1-Terminus。
At present, official Apps, webends, and applets are being updated simultaneously to DeepSeek-V3.2-Exp。
This update is here API Large price reductions will reduce the developer 's cost of calling DeepSeek API by 50% or more.
While the validity of DeepSeek-V3.2-Exp has been validated on the public assessment set, it still requires a wider and larger-scale test in the real-life scene of users. Additional API access interfaces have been temporarily reserved for V3.1-Terminus for user-friendly comparison tests。
DeepSeek-V3.2-Exp Model is now in Huggingface with magicOpen Source:
- HuggingFace:
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
- ModelScope:
https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Exp
The papers have been published simultaneously:
https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf