This episode describes a new multimodal artificial intelligence model called 'MoMa' developed by Meta. MoMa is based on an 'early fusion' architecture that combines text and images into a single model. The article highlights MoMa's efficiency, demonstrating how it significantly reduces computational cost through the use of 'sparse modality-aware' techniques, which leverage 'mixture-of-experts' (MoE) and 'mixture-of-depths' (MoD) to optimize the use of computational resources. Additionally, the article explores the application of 'upcycling' to improve the model's performance. The research conducted experiments on various MoMa models, evaluating their performance and throughput, and identified the optimal architecture for different tasks. The article concludes with a discussion of MoMa's current limitations and the promising research directions for future developments.