In recent years, transformer-based visual tracking models have shown significant advancements by leveraging global feature representations to improve information interaction. However, their high computational demands hinder efficient deployment on resource-constrained platforms like UAVs and robotic devices. To address this, we propose OneStar, a template-branch separable vision transformer that balances efficiency and accuracy. Unlike one-stream trackers that process the template in every frame, OneStar performs template feature extraction and fusion only during initialization, reducing redundant computations. A tracking token is designed to guide the use of prior positional and size information from the initial frame, enhancing fusion accuracy. We also provide two model variants optimized for different edge devices, ensuring efficient resource use while maintaining strong tracking performance. Extensive evaluations show that OneStar delivers competitive accuracy with state-of-the-art models while achieving excellent real-time performance on edge platforms. For example, OneStar achieves a 4.9% AO improvement on GOT-10k compared to other real-time trackers.