UI-TARS is ByteDance’s native GUI agent model: a multimodal model trained to perceive screenshots, reason about interfaces and operate desktop, mobile and web applications directly โ clicking, typing and scrolling like a person. The open release has earned about 11,600 stars on GitHub.
Most agent systems glue a general vision model to external grounding and planning modules. UI-TARS instead trains a single native model for the whole loop: unified perception of screenshots, precise grounding of UI elements from natural-language instructions, and multi-step action execution with reasoning about what to do next. The repository releases the UI-TARS model family (built at 7B and larger scales on a vision-language backbone) together with inference code, while companion work from the team covers the desktop agent application built on the model. The accompanying research describes capabilities such as reflective ‘System-2’ style reasoning over action trajectories and online learning from experience, and the models posted strong results on GUI benchmarks such as ScreenSpot and OSWorld-style evaluations at release. Because the weights are open, researchers and product teams can fine-tune or self-host a GUI agent instead of depending on a proprietary computer-use API. The repository has remained active into 2026.
UI-TARS is free and open source under the Apache-2.0 license, with model weights released publicly. Your only costs are the hardware or cloud GPUs you run it on.
UI-TARS is for researchers and product teams building computer-use agents โ desktop automation, software testing, accessibility tools โ who want an open, specialized GUI model they can inspect, fine-tune and deploy on their own infrastructure.