ByteDance has developed UI-TARS, a new AI system that can autonomously control computers and mobile devices to perform complex tasks. According to research published on Arxiv and reported by Taryn Plumb for VentureBeat, the system outperforms existing AI models like GPT-4o and Claude across multiple benchmarks. UI-TARS uses both 7B and 72B parameter versions and was trained on approximately 50 billion tokens. The system can understand graphical interfaces, execute multi-step tasks, and explain its reasoning process in real-time. It works across desktop, mobile, and web applications, using a combination of text, images, and interactions to navigate various environments. The researchers equipped UI-TARS with both short-term and long-term memory capabilities, allowing it to learn from mistakes and adapt to new situations.