Multimodal Visual Question Answering (VQA) have gained a lot of attention because of their ability of solving complex problems like puzzle related problems by integrating visual and textual data provided by user. Traditional LLMs like ChatGPT and Gemini though powerful in language processing, struggle with tasks requiring deep understanding of visual information and algorithmic reasoning. This research paper aims to bridge the gap by developing model capable of interpreting multimodal data enabling accurate puzzle solving. Traditional Convolutional Neural Network (CNN) + Long Short Term Memory (LSTM) architecture is modified by utilizing attention mechanism and embeddings. AlgoPuzzleVQA dataset is used to train the proposed model which consist of 18 different types of algorithmic puzzles(classes) involving 100 instances associated with each puzzle, hence totaling to 1800 instances. Proposed model achieves average accuracy of 44.4% after training for 50 epochs and also other metrics like training loss, validation loss is marginal. The proposed model outperforms traditional LLMs in 8 out of 18 puzzle types and achieves maximum accuracy of 100% in wood slide puzzle problem. Multimodal VQA has many real-time applications, including assisting visually impaired people through automatic VQA systems, enhancing medical diagnosis by answering queries related to medical reports, and improving educational platforms by providing AI-driven tutoring support. All these applications showcase the real-world problem-solving capability of multimodal VQA and make it a valuable advancement in AI research. The focus of this paper is to improve the algorithmic reasoning ability of LLMs by developing a model that will bridge this gap.