UI-TARS-1.5

A model that operates screens

UI-TARS-1.5 is a model built to work a graphical user interface: it looks at what is on the screen and decides what to click, type or scroll next. ByteDance Seed released it on 16 April 2025 as the successor to UI-TARS.^[1] Two versions were announced at the same time. The smaller UI-TARS-1.5-7B was open-sourced, while the top-performing model was not offered for direct download and was instead made available through research access requests.^[1]

What is open

The open release is published under the Apache-2.0 license.^[1] It is small relative to earlier work in the same line, but it already outperforms a much larger predecessor: UI-TARS-1.5-7B scored 27.5 on the OSWorld benchmark, above the original UI-TARS-72B-DPO's 24.6.^[1] That comparison is the clearest single statement about the generation: the same level of screen competence arrives in a model far cheaper to run.

Benchmark results

On OSWorld, the full model reported 42.5 at 100 steps. For comparison, OpenAI CUA reported 36.4 and Claude 3.7 reported 28, while the previous best figure on the benchmark was 38.1, reached at 200 steps.^[1] The step count matters because an agent that needs fewer interactions to finish a task is also less exposed to compounding errors.

On AndroidWorld the model reported 64.2 against a previous best of 59.5. On the screen-understanding benchmarks it reported 94.2 on ScreenSpot-V2 and 61.6 on ScreenSpot-Pro, where OpenAI CUA reported 23.4 and Claude 3.7 reported 27.7.^[1] The ByteDance Seed blog puts the previous best on ScreenSpot-Pro at 43.6%.^[2]

ByteDance Seed says the model reaches state of the art on 7 GUI benchmarks.^[2] The ScreenSpot-Pro gap is the most striking of the published numbers: the benchmark uses small, dense interface elements where general-purpose vision models tend to struggle, and the reported margin over both OpenAI CUA and Claude 3.7 is wide.^[1]

Games as a test bed

Beyond office-style interfaces, the Seed blog reports that UI-TARS-1.5 achieved the highest success rate on both the Mine Blocks and Kill Mobs tasks in Minecraft, tested against OpenAI VPT, DeepMind DreamerV3 and JARVIS-VLA.^[2] The team also tested it on 14 small games from poki.com, with a limit of 1,000 steps per game.^[2] Games serve as a stress test for long sequences of actions where the screen changes constantly and there is no fixed list of buttons to press.

Where it sits

UI-TARS-1.5 arrived at a moment when several labs were shipping models that drive computers rather than answer questions. Its significance is twofold. First, the reported numbers place it ahead of contemporaneous general-purpose computer-use models on OSWorld, AndroidWorld and the ScreenSpot benchmarks.^[1] Second, and more durable for research, one version is downloadable under a permissive license, so groups without access to a frontier lab's internals can run, evaluate and build on a competitive screen agent.^[1] The split release — an open small model and a stronger gated one — reflects a pattern that became common for agent models: publish enough to be useful, keep the best weights behind a request form.^[1]

What the benchmarks do not settle is reliability over long, messy real-world tasks. A reported 42.5 at 100 steps on OSWorld is progress on a hard benchmark, not a claim that the model can finish arbitrary work on a desktop unsupervised.^[1]