With the fundamentals of MoE architecture and distributed training established, the focus now shifts to preparing these models for real-world application through inference optimization and deployment. While sparsity offers advantages in training compute, it introduces unique considerations for efficient inference, such as managing latency and memory effectively. This chapter covers strategies to address these issues. We will examine techniques including specialized batching approaches, model compression methods adapted for MoE structures, leveraging hardware acceleration, optimizing router behavior at inference time, and suitable deployment patterns for large sparse models.