Aguvis-72B
Salesforce AI Research · 2024-12-05
Aguvis-72B is a paper-reported 72B-class pure-vision GUI agent developed by Salesforce AI Research and the University of Hong Kong for autonomous cross-platform graphical user interface interaction. It operates directly on screen images using a unified action space and inner-monologue reasoning across mobile, desktop, and web environments. The paper evaluates this model, but the public Aguvis repository links only a 7B checkpoint and still marks further weights as forthcoming, so this row represents the evaluated 72B model rather than a downloadable artifact.
Benchmark scores
| Benchmark | Score |
|---|---|
| OSWorld | 10.3 |