The integration of different modalities in deep learning models facilitates the incorporation of various data forms such as images and text, which enhances the multi-modal task performance. These models tackle issues such as representation of features, alignment of modalities, and strategies for fusion. The state of the art utilizes contrastive architectures such as CLIP, ALIGN, vision-language transformers like ViLT and Flamingo, and other hybrid components to boost cross-Moden reasoning. Tasks include image captioning, visual question answering, and multi-modal retrieval. Subsequent objectives combine architectures with efficiency in training and alignment techniques. Multi-modal learning is instrumental in pushing the boundaries of AI and its applications in comprehending the multifaceted nature of the real world.