Cua Releases Small Specialized Model 'CUA-S1' and Open-Source Ecosystem for Computer-Use Agents

The Cua ecosystem has released new open-source components and the model CUA-S1 for developing Computer-Use agents. This infrastructure supports the 'Computer-Use 2.0' concept, which involves navigating between code, APIs, and GUIs within a single task.

CUA-S1 Model Features

• System 1 Model: A small model specialized for fast, bounded decisions such as form input, rather than complex planning or reasoning.
• How it Works: Instead of generating response tokens, it scores decisions based on structured interface elements and document values; application code controls the action sequence, and Cua Driver handles the execution boundary.
• Release Scope: Python model code, synthetic data generation, training, and evaluation code are released on GitHub under the MIT license, while model weights are hosted separately on Hugging Face.

Key Open-Source Tools

• Cua Driver: A tool that allows agents to inspect and manipulate native apps and browsers on macOS, Windows, and Linux, supporting CLI, MCP, and typed SDKs.
• Cua Fleets: Provides isolated cloud desktop environments, enabling command execution and screenshot capture via the Sandbox SDK.
• Lume: Creates and manages local macOS/Linux VMs on Apple Silicon using Virtualization.Framework.
• Cua Bench: Offers features for building computer-use tasks, evaluating agents, and exporting training trajectories. It allows starting with simulated tasks without needing VMs or model API keys.