os-atlas.md

OS-Atlas

Overview

OS-Atlas is an open-source foundation action model for generalist GUI agents, developed by Shanghai AI Lab. The project is officially hosted and described in their research paper.

Key Features

• Multiple model sizes (4B and 7B)
• GPT-4o integration option
• Cross-platform support (mobile, desktop, web)
• Large-scale GUI grounding corpus (13M+ elements)
• Developed by Shanghai AI Lab

Performance

OSWorld Results

• OS-Atlas-Base-7B w/ GPT-4o: 14.63%
• OS-Atlas-Base-4B w/ GPT-4o: 11.65%

ScreenSpot Results

• OS-Atlas-7B: 82.5% accuracy
• OS-Atlas-4B: 71.9% accuracy

Technical Details

• Model Sizes: 4B and 7B parameters
• Optional GPT-4o integration
• Focus: GUI grounding and interaction
• Supports multiple operating systems
• Execution-based evaluation

References

• Project: https://osatlas.github.io/
• Paper: https://arxiv.org/abs/2410.23218
• Code: https://github.com/OS-Copilot/OS-Atlas
• Models: https://huggingface.co/OS-Copilot/OS-Atlas-Action-7B