Multimodal Large Language Models (MLLMs) demonstrate strong reasoning capabilities by integrating visual and textual information with external knowledge, but their computational demands limit practical deployment. We propose a knowledge-guided structured pruning approach that leverages external knowledge graphs to inform compression decisions. Our method achieves a favorable trade-off between model size and performance: at 30% compression, we retain 89.2% of original accuracy while reducing inference time by 1.4x. Experiments on knowledge-grounded visual question answering show modest improvements over magnitude-based pruning baselines, though we observe increased hallucination rates typical of compressed models. Our approach provides a practical framework for deploying MLLMs in resource-constrained environments while maintaining reasonable performance on knowledge-intensive tasks.