Developed and integrated a new Mixture of Experts (MoE) switch model into the apple/axlearn repository, targeting scalable experimentation on TPU v6e and Fuji architectures. Leveraging Python and deep learning frameworks, the work introduced architecture enhancements and utilities to infer optimal batch sizes from mesh shapes, enabling efficient distribution and improved throughput. Expanded the test suite to comprehensively validate both the MoE switch model and Fuji-specific configurations, including rematerialization-aware training setups for better memory and performance optimization. This engineering effort established a foundation for production-ready, cost-efficient MoE workflows on advanced TPU hardware, emphasizing model optimization and robust machine learning practices.