With the rise of short video platforms (e.g., TikTok), the forms of news have shifted from text-images to short videos. This shift introduces new challenges due to the complexity of more modalities. Consequently, developing new methods for fake news video detection has become urgent. However, previous methods have not comprehensively and systematically considered the discrepancy of modal contributions and feature distribution. The discrepancies not only reduces the performance of the fusion network but also suppresses the contributions of nondominant modalities, leading to suboptimal results. To address this, we propose a fake news video detection method called AMFNVD, which is Adaptive Multi-gate Mixture-of-Experts(AMMoE) and Multi-branches joint loss for Fake News Video Detection. AMMoE we presented is to align and refine to mitigate the problem of feature distribution discrepancy. Multi-branches joint loss balance modal contributions. Extensive experiments on two fake news datasets demonstrate that AMFNVD outperforms state-of-the-art methods. Our codes are available in https://anonymous.4open.science/r/AMFNVD-main.