UI-TARS-1.5

UI-TARS-1.5 is ByteDance's open vision-language-action model for GUI automation, a native agent model trained to perceive screens and act on them. It works from screenshots alone across desktop, mobile, and web, and the 1.5 generation adds reinforcement-learning-driven System-2 reasoning: thinking through steps, reflecting on mistakes, and correcting course, which lifted it to state-of-the-art results on OSWorld, ScreenSpot-Pro, and game benchmarks at release.

The UI-TARS-1.5-7B weights were open-sourced in April 2025 under Apache 2.0, with short-term and long-term memory mechanisms for multi-step tasks and an SDK plus desktop app in the ecosystem. Caveats: ByteDance has since released UI-TARS-2 with broader game, code, and tool-use abilities, so 1.5 is now the previous generation, and real-world desktop automation still needs guardrails around destructive actions.