ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data

Benchmark Model Rank Results
computer-use-agents-on-windowsagentarena-15-stepsScaleCUA-32B#4Success Rate: 21.4
computer-use-agents-on-osworld-50-stepsScaleCUA-32B#6Success Rate: 17.7
computer-use-agents-on-mmb-gui-mmbench-guiScaleCUA-32B#8Accuracy: 82.0
computer-use-agents-on-osworld-15-stepsScaleCUA-32B#9Success Rate: 16.5
computer-use-agents-on-windowsagentarena-50-stepsScaleCUA-32B (50 steps)#11Success Rate: 24.2
computer-use-agents-on-osw-g-osworld-gScaleCUA-32B#12Accuracy: 60.6
computer-use-agents-on-screenspot-v2ScaleCUA-32B#12Accuracy: 94.7
computer-use-agents-on-screenspot-proScaleCUA-32B#23Accuracy: 59.2
computer-use-agents-on-androidworldScaleCUA-32B#31Success Rate: 30.6