UI-TARS-2

UI-TARS-2 is a GUI agent from the UI-TARS line. Its technical report was submitted to arXiv on 2 September 2025, and the release was announced on GitHub on 4 September 2025 as a major upgrade from UI-TARS-1.5 with an “all in one” agent design.^[1][2] GUI stands for graphical user interface: a GUI agent is meant to operate screens, not only answer text.

Training and model stack The report says UI-TARS-2 was trained with multi-turn reinforcement learning for GUI agents.^[1] In reinforcement learning, a system learns from feedback on its actions rather than only from labeled examples; multi-turn means the feedback covers a sequence of steps. For GUI work this matters because an agent often has to act, observe the result and act again over several turns. The model is initialized from the pre-trained checkpoint of Seed-thinking-1.6, which has a 532M-parameter vision encoder and a mixture-of-experts LLM with 23B active and 230B total parameters.^[3] The technical report lists Haoming Wang and 111 other authors, 112 in total.^[1] A second version of the arXiv report was revised on 5 September 2025.^[1]

Reported scores The report gives 88.2 on Online-Mind2Web, 47.5 on OSWorld, 50.6 on WindowsAgentArena and 73.3 on AndroidWorld.^[1] Across a suite of 15 games it reached a mean normalized score of 59.8, about 60% of human level, and was competitive with OpenAI o3 on LMGame-Bench.^[1] The authors said the GUI results beat strong baselines such as Claude and OpenAI agents.^[1] These are separate benchmark results, not a single aggregate; each covers a different kind of environment or task.

GUI-only and GUI-SDK settings The report separates a GUI-only setting from a GUI-SDK setting. In the GUI-SDK setting, UI-TARS-2 scored 45.3% on Terminal Bench and 68.7% on SWE-Bench.^[3] On BrowseComp-zh it scored 32.1% in the GUI-only setting and 50.5% in the GUI-SDK setting; on BrowseComp-en the two figures were 7.0% and 29.6%.^[3] The report lists both settings for these two benchmarks, so the difference between them can be read directly from the numbers.

Place in the line UI-TARS-2 is presented as a step beyond UI-TARS-1.5: the later model keeps the agent framing and adds multi-turn reinforcement learning, with tests that cover web, desktop, Android, games, terminal and software-engineering benchmarks.^[1][3] The report also records the model stack, naming a large vision encoder and a mixture-of-experts language model as its starting point.^[3] Its place in the line is that of an all-in-one agent evaluated across several GUI environments, with the report’s own scores and baseline comparisons as the evidence.