Longhui Zhang implemented support for the JoyAI LLM-Flash model on NPU devices within the jd-opensource/xllm repository, focusing on optimizing performance for specialized hardware. Using C++ and leveraging expertise in NPU optimization, deep learning, and machine learning, Longhui tailored weight merging and tensor operations to the NPU architecture. This work enhanced hardware compatibility and enabled more efficient inference, addressing the need for broader deployment of large language models on NPU-accelerated platforms. The integration aligned with hardware team objectives by reducing deployment friction for customers and improving resource utilization, demonstrating a focused and technically deep approach to model optimization.