SOTA on mobile agent tasks, surpassing human baseline of 80%.

First vision-language model to lead AndroidWorld. Beats Claude Opus 4.6 (62.0) by 13.7 points on Android GUI tasks.

Mobile GUI agent. On par with UI-TARS-2 (73.3) at only 8B params.

Seed1.8 achieves 61.9% on OSWorld and 85.9% on OnlineMind2Web, surpassing Claude Sonnet 4.5 and GPT-O3-CUA on computer use tasks. Also achieves 70.7% on AndroidWorld.

New SOTA for open models on AndroidWorld. Best sub-30B agent model.

Strong mobile performance. Multi-platform fundamental GUI agent.

Mobile agent benchmark. Competitive with specialized agent frameworks.

VeriGUI-7B achieves best open-source results on AndroidControl High (74.2% TM, 65.5% GR, 51.1% SR) and competitive with GPT-5.1 and Gemini-3-flash. Introduces TVAE framework (Thinking-Verification-Act