UI-TARS
End-to-end open GUI agent that sees only screenshots and acts with keyboard and mouse. It scores 24.6 on OSWorld (50 steps) vs Claude's 22.0.
Native GUI-agent model (2B, 7B, 72B) trained on screenshot perception, unified action space, System-2 reasoning and iterative self-improvement. UI-TARS-1.5 followed in April 2025, and UI-TARS-2 (arXiv 2509.02544) in September.
- Date
- Tuesday, 21 January 2025
- Lab
- ByteDance Seed
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| OSWorld (50 steps) | 24.6 vs Claude 22.0 | company |
| OSWorld (15 steps) | 22.7 vs Claude 14.9 | company |
| AndroidWorld | 46.6 vs GPT-4o 34.5 | company |
arXiv v1 2025-01-21; HF weights uploaded 2025-01-22 UTC. Apache-2.0 per HF card. Benchmarks are authors' own.
Sources
This record was checked against its sources on 6 October 2026. How we check