extreme1228/ScaleCUA-qwen3.5-osworld
The extreme1228/ScaleCUA-qwen3.5-osworld is a 9 billion parameter Qwen3.5-based computer-use agent developed by ScaleCUA, specifically trained for the OSWorld benchmark. This model utilizes online Reinforcement Learning (RLVR) and achieves an open-source state-of-the-art performance of 68.7% on OSWorld. It is optimized for complex computer interaction tasks through verifiable GUI-task synthesis and efficient online RL techniques.
Loading preview...
Overview
This model, extreme1228/ScaleCUA-qwen3.5-osworld, is a 9 billion parameter computer-use agent built upon the Qwen3.5 architecture. Developed by ScaleCUA, it is specifically designed and trained for the OSWorld benchmark, focusing on enabling agents to interact with computer interfaces. The model represents the online-RL checkpoint of the ScaleCUA project, which leverages advanced techniques like VeriGen (verifiable GUI-task synthesis), Frontier Sampling, and Visual Context Segmentation to scale online Reinforcement Learning for computer-use agents.
Key Capabilities
- Computer-Use Agent: Functions as an agent capable of interacting with computer environments.
- OSWorld Optimization: Specifically trained and optimized for performance on the OSWorld benchmark.
- Online Reinforcement Learning (RLVR): Utilizes an online RL stage for enhanced performance.
- State-of-the-Art Performance: Achieves 68.7% on OSWorld, marking it as an open-source state-of-the-art result.
- Advanced Training Techniques: Incorporates VeriGen for verifiable GUI-task synthesis and efficient online RL methods.
Use Cases
This model is primarily intended for research and development in the field of computer-use agents and automated GUI interaction. It is particularly suitable for:
- Benchmarking: Evaluating and advancing performance on the OSWorld benchmark.
- Agent Development: Serving as a foundation or comparison point for building new computer-use agents.
- Research: Exploring techniques in online reinforcement learning, GUI-task synthesis, and visual context understanding for agent training.