Document-level Relation Extraction aims to identify relationships between entity pairs in long texts, which is crucial for building knowledge graphs. However, it faces long-tail problem and evidence focus problem. To address these issues, this paper proposes the ATLOP-DME (De-Sampling and Mixture of Experts) model, building on the ATLOP method. The model introduces two core components: 1) De-Sampling to balance the relation category distribution and alleviate the long-tail problem, and 2) a Mixture of Experts system to focus on evidence. We conducted training on the DocRED, Re-DocRED, CDR, and GDA datasets. The experimental results demonstrate that, compared to the ATLOP model, the ATLOP-DME model achieved improvements of 24.02%, 10.19%, 7.5%, and 2.9% on the respective datasets. Notably, it sets new SOTA performance on the DocRED and Re-DocRED datasets. The code is now publicly available on GitHub ( https://github.com/Shenjingbang/ATLOP-DME ).