People are increasingly using social media platforms to connect, share information, promote ideas and products, express opinions, and engage in common interests. However, this surge in social media usage has also resulted in an increase in hate speech and aggressive language on these platforms. The existence of harmful content on social media platforms is highly undesirable as it significantly hinders positive and secure social interactions online. This study developed a multimodal model for detecting abusive language by combining convolutional techniques with fine-tuned BERT (Bidirectional Encoder Representations from Transformers). The novelty of this approach lies in integrating text features with image features using CNN and Fine-tuned BERT. Pre-trained CNN models such as ResNet-50, VGG19, and Inception-V3 were utilized for extracting image features, while fine-tuned BERT was employed for extracting text features in detecting abusive language across multiple modalities in Amharic. The results of the multimodal experiment demonstrate that the model combining CNN and Fine-Tuned BERT achieved superior performance, achieving an accuracy of 75% in detecting abusive language in Amharic.