ByteDance releases UI-TARS-1.5 and open-sources its 7B model

Background

UI-TARS is ByteDance Seed's family of GUI agents, which operate software by looking at screenshots and using the keyboard and mouse. UI-TARS-1.5 is the update to the original model, released on 16 April 2025.^[1]

What happened

On 16 April 2025 ByteDance Seed shared UI-TARS-1.5 and open-sourced UI-TARS-1.5-7B. The 7B means the open model has about 7 billion parameters, the adjustable numbers that a model learns during training.^[1]

The full UI-TARS-1.5 model was offered through research access requests.^[1]

Key results

The full model scored 42.5 on OSWorld with 100 steps. That was ahead of OpenAI CUA at 36.4 and Claude 3.7 at 28.^[1] On AndroidWorld it scored 64.2, and on ScreenSpot-Pro, a benchmark for locating interface elements in high-resolution screens, it scored 61.6.^[1]

The open 7B model scored 27.5 on OSWorld.^[1]

Open license

The release is under the Apache-2.0 license, which allows others to use, modify and redistribute the model with the license notice kept.^[1]

Open weights matter to researchers because they can be inspected, run on local hardware and modified for experiments. A model offered only through an application process is useful for published benchmarks, but it is harder for an outside team to reproduce the results. The number of parameters is one of the main factors in how much memory and computing a model needs to run, so a 7B model can run on far more modest hardware than a much larger one.

A parameter is a single adjustable value inside the network, and the 7B figure gives the number of them in the open model. Release decisions for models like this often balance openness and control: a smaller model can be shared widely while the largest version is kept for approved research use.

Significance

The release gave researchers and developers an open model they could run and study themselves, while the full model was available only by application.