ShowUI

End-to-end vision-language-action model for GUI agents (CVPR 2025): a 2B model (built on Qwen2-VL-2B-Instruct) with UI-guided visual-token selection to cut redundant screenshot tokens and interleaved vision-language-action streaming for navigation — ShowUI-2B is ShowLab's most-downloaded HuggingFace model and a common base for open computer-use experiments (1.9K★).

Joint ShowLab / Microsoft work, NUS-led. The line continues with ShowUI-π (CVPR 2026), a flow-based generative policy for continuous GUI manipulation, and ShowUI-Aloha (2026), a human-taught desktop computer-use agent.