UI-TARS 72B SFT
ByteDance Seed · 2025-01-20 · 73.4B parameters
UI-TARS 72B SFT is the largest supervised-fine-tuned checkpoint in ByteDance Seed's original UI-TARS family, produced from a Qwen2-VL-72B backbone by continual pre-training and an annealing phase over GUI perception, grounding, action-trace, and reflection-tuning data. It is an end-to-end vision-language GUI agent that takes task text and screenshots and generates structured reasoning and actions for desktop, web, and mobile interfaces. The paper evaluates both direct-action system-1 and thought-before-action system-2 prompting for UI-TARS-72B, so reasoning is switchable at inference rather than unknown.
Benchmark scores
| Benchmark | Score |
|---|---|
| OSWorld | 18.8 |