UI-TARS 7B SFT
ByteDance Seed · 2025-01-20 · 8.3B parameters
UI-TARS 7B SFT is ByteDance Seed's open-weight, 7B-scale SFT GUI agent checkpoint obtained after continual pre-training and annealing of Qwen-2-VL-7B on an approximately 50-billion-token GUI-focused data mixture spanning perception, grounding, action traces, reasoning, and reflection. It consumes text and screenshots and generates text-form thoughts and unified control actions for desktop, web, and mobile tasks. The paper states that UI-TARS-7B supports prompt-selected no-thought (System 1) and with-thought (System 2) inference, and this SFT checkpoint is the reference policy for the subsequent DPO phase; the exact cumulative training-token count is not disclosed.
Benchmark scores
| Benchmark | Score |
|---|---|
| OSWorld | 17.7 |