UI-TARS: Pioneering Automated GUI Interaction with Native Agents

Benchmark Model Rank Results
computer-use-agents-on-mind2web-cross-domainUI-TARS-72B-SFT#1Element Accuracy: 68.9Step Success Rate: 62.1
computer-use-agents-on-mind2web-cross-taskUI-TARS-72B-SFT#1Element Accuracy: 74.7Step Success Rate: 68.6
computer-use-agents-on-mind2web-cross-websiteUI-TARS-72B-SFT#1Element Accuracy: 72.4Step Success Rate: 63.5Operation F1: 91.2
computer-use-agents-on-online-mind2webUI-TARS-1.5 (early access)#2Success Rate: 75.8
computer-use-agents-on-osworld-15-stepsUI-TARS-72B-DPO#4Success Rate: 22.7
computer-use-agents-on-osworld-15-stepsUI-TARS-7B-DPO#7Success Rate: 18.7
computer-use-agents-on-webvoyagerUI-TARS-1.5 (early access)#4Success Rate: 84.8
computer-use-agents-on-osworld-50-stepsUI-TARS-72B-DPO#5Success Rate: 24.6
computer-use-agents-on-windowsagentarena-50-stepsUI-TARS-1.5 (early access)#6Success Rate: 42.1
natural-language-visual-grounding-on-screenspotUI-TARS-7B-SFT#7Accuracy (%): 89.5
computer-use-agents-on-screenspot-proUI-TARS-1.5 (early access)#14Accuracy: 61.6
computer-use-agents-on-screenspot-v2UI-TARS-1.5 (early access)#14Accuracy: 94.2
computer-use-agents-on-androidworldUI-TARS-1.5 (early access)#18Success Rate: 64.2