Meningiomas are among the most common primary intracranial tumors in adults, exhibiting varying biological behaviors across different grades. High-grade meningiomas display aggressive behavior, and the survival and prognosis of patients largely depend on the treatment strategy. Therefore, preoperative grading for meningiomas holds significant clinical importance. In recent years, computer-aided diagnosis techniques based on radiomics and deep learning methods have achieved notable success due to their non-invasive advantages for preoperative evaluation. However, most studies face limitations due to the singular source of dataset information, resulting in lower accuracy and reliability issues. To solve the above issues, we design a multimodel classification network named 3D-MGNet to significantly enhance the preoperative grading of meningiomas by integrating dual MRI modalities: T1 CE and T2 Flair. The proposed network consists of a Dual Multimodal Meningioma Feature Extractor (DME) designed to leverage the distinct modalities, and a deep-radiomics features integration discriminator for robust meningiomas grading. A key innovation within our model is the 3D Large Kernel Attention (3D-LKA) mechanism, which efficiently captures global and local characteristics from processes complex, memory-intensive 3D data. This efficiency attributes to the amalgamation of spatial local convolution operations, spatial long-distance convolution operations, and channel convolution operations. Extensive experimentation demonstrates superior performance of the proposed 3D-MGNet in meningioma grading accuracy compared to existing models.