We deployed the hand tracking model on different platforms: IOS, Android, PC. To leverage the best performance of the devices, we adjusted the model to adapt to the specific harewares. For example, we converted the model to CoreML to exploit the Neural Engine rather than the GPU on IOS device. We quantized the model in order to utilize Acceralorator to run the inference on the Android device. In addition, on the PC platform, we fitted a pre-crafted 3D hand model to match up the detected 2D landmarks via Ceres-Solver.