CUA-Lite Releases Open-Source Tools for Local Model Computer Use

CUA-Lite is an open-source framework for developing and training Computer Use Agents (CUA) that help models directly manipulate desktop apps, web forms, and mobile apps. It provides all four core components—model harness, sandbox, data, and training framework—as open source.

Key Components

• Harness: A modular structure responsible for managing screenshots and action history, and converting clicks and keyboard inputs. It supports 14 series of open-weight models, including GPT, Claude, Gemini API models, Qwen3-VL, and UI-TARS, allowing easy addition of models to existing inference stacks.
• Sandbox: A Docker-based sandbox that operates without VMs, containing over 30,000 CUA tasks with verifiable reward signals. It reproduces benchmarks and training environments such as OSWorld, CUAGym, and CUAWorld, and can run in environments where VMs are not possible, such as WSL and cloud instances.
• Data: Over 10 SFT datasets and the latest rollouts (interaction trajectories) from frontier CUAs are provided for free in a single format on Hugging Face.

Execution and Training Framework

Through a single agent-environment interface, various agents, including local models, can be connected to the sandbox. The generated rollouts are reused for the following three purposes:

• Eval: Integrates over 15 benchmarks, such as OSWorld, WebArena, and AndroidWorld, to rank agent performance.
• SFT: Renders the single trajectory format to fit each model adapter, enabling cross-model training, such as fine-tuning Qwen3-VL with rollouts from GPT-5.5.
• RL: Performs reinforcement learning by utilizing them as training signals.