Models

← All models

UI-TARS 72B DPO

ByteDance Seed · 2025-01-22 · 73.4B parameters

UI-TARS 72B DPO is ByteDance Seed's largest original UI-TARS checkpoint, a dense Qwen2-VL-based native GUI agent that accepts text instructions and screenshots and produces text-form thoughts and human-like keyboard and mouse actions. Its third training phase applies Direct Preference Optimization to reflective corrected-versus-erroneous action pairs, improving action selection for long-horizon GUI tasks across desktop, web, and mobile environments. The paper reports an approximately 50-billion-token training-data size across the three phases, so exact cumulative training exposure is undisclosed.

Benchmark scores