UI-TARS 7B DPO
ByteDance Seed · 2025-01-22 · 8.3B parameters
UI-TARS 7B DPO is the DPO-finalized, 7B-class checkpoint of ByteDance Seed's original UI-TARS native GUI agent, built by continually training Qwen2-VL and then optimizing corrected actions over erroneous ones. It consumes task text and GUI screenshots, emits thought-and-action text for end-to-end control, and uses a unified action space across web, desktop, and mobile environments.
Benchmark scores
| Benchmark | Score |
|---|---|
| OSWorld | 18.7 |