Yiping Ma developed a new Mixture of Experts (MoE) switch model for the apple/axlearn repository, targeting scalable experimentation on TPU v6e and Fuji architectures. Using Python and leveraging deep learning and model optimization techniques, Yiping enhanced the model architecture and introduced utilities to infer batch sizes from mesh shapes, enabling efficient distribution and improved throughput. The work included comprehensive test coverage to validate both the MoE switch model and Fuji-specific configurations, such as rematerialization-aware training for optimized memory usage. This engineering effort established a robust foundation for large-scale, cost-efficient MoE experimentation and production readiness on advanced TPU platforms.