OpenGVLab/ScaleCUA-3B

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Sep 16, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ScaleCUA-3B by OpenGVLab is a 3 billion parameter vision-language model designed for cross-platform computer use automation, capable of interacting with graphical user interfaces across 6 operating systems and 3 task domains. It is built on a large-scale, diverse corpus collected via a novel dual-loop interactive pipeline. The model excels at GUI understanding, grounding, and task completion, achieving significant improvements on benchmarks like WebArena-Lite-v2 (+26.6 points) and ScreenSpot-Pro (+10.7 points), making it suitable for developing general-purpose computer-use agents.

Loading preview...

Overview

ScaleCUA-3B is a 3 billion parameter vision-language model developed by OpenGVLab, specifically designed for automating interactions with graphical user interfaces (GUIs) across diverse platforms. It is built upon a unique, large-scale dataset collected through a dual-loop interactive pipeline involving both automated agents and human experts. This dataset spans 6 operating systems and 3 task domains, addressing the limitations of existing open-source datasets for cross-platform general-purpose computer-use agents.

Key Capabilities

  • Cross-Platform GUI Automation: Seamlessly operates across heterogeneous platforms, including desktops, mobile devices, and web browsers.
  • Enhanced Performance: Demonstrates consistent gains on various benchmarks, improving absolute success rates by +26.6 points on WebArena-Lite-v2 and +10.7 points on ScreenSpot-Pro compared to baselines.
  • State-of-the-Art Results: Achieves high performance on GUI-related benchmarks, including 94.4% on MMBench-GUI L1-Hard, 60.6% on OSWorld-G, and 47.4% on WebArena-Lite-v2.
  • Dual Interaction Modes: Supports a Direct Action Mode for immediate GUI grounding and a Reasoned Action Mode for complex, multi-step tasks requiring a thinking process before action generation.

Good for

  • Developing autonomous agents for computer use automation.
  • Tasks requiring precise GUI understanding and interaction across different operating systems.
  • Research and development in vision-language models for agentic workflows.
  • Applications needing robust GUI grounding and reasoned action generation.

Finetunes
1 models