ByteDance open-sources the UI-TARS GUI agent model

Background

A GUI agent is an AI system that uses software through the graphical interface, the same windows, buttons and menus a person sees. A system that works from screenshots alone can, in principle, be used on any screen a person can see.

What happened

On 21 January 2025 ByteDance Seed researchers posted the UI-TARS paper to arXiv.^[1] UI-TARS is what the authors call a native GUI agent. It perceives only screenshots and acts through keyboard and mouse operations.^[1]

Key results

The paper reported that UI-TARS scored 24.6 on OSWorld with 50 steps, compared with 22.0 for Claude. OSWorld is a benchmark of tasks on a real desktop operating system. On AndroidWorld it scored 46.6, against 34.5 for GPT-4o.^[1]

The authors reported state-of-the-art results on more than 10 GUI-agent benchmarks.^[1]

Open release

The model weights were released under the Apache-2.0 license on GitHub, including UI-TARS-72B-DPO.^[2] The open license means that others may use, modify and redistribute the model, provided they keep the license notice.

Follow-up versions

UI-TARS-1.5 followed on 16 April 2025 and UI-TARS-2 on 4 September 2025.^[2] Those releases are covered in their own entries.

Benchmarks for GUI agents differ in what they test. Some run on a desktop operating system, some on a phone system and some in a web browser. Scores are usually the share of tasks completed on a fixed set, so results from different papers can be compared only when the test and the step limit are the same.

Working from screenshots alone means the agent has to recognise buttons, text and layout from the pixels on the screen. That is a harder perception problem than reading a structured list of interface elements.

Significance

The paper's comparisons with Claude and GPT-4o on standard tests are the main evidence for the approach. The open weights also let other researchers study and test this kind of agent directly.