Cua releases 2.8 MB specialist AI model for web form automation

Startup Cua has open-sourced CUA-S1-FORMS, a 706,000-parameter model designed to handle web form interactions without relying on large general-purpose language models. The model scores possible actions — such as clicking, checking, or skipping form fields — rather than generating text token by token, making it highly focused on narrow decision-making. In internal evaluations, the specialist model achieved 99.7% accuracy on form decisions compared to 83.6% for a hosted general model, while also running locally in 7–9 milliseconds versus 260–280 milliseconds for cloud-based alternatives. The model was trained on synthetic data in under 30 minutes and is released under an MIT license. Cua says the project explores a middle ground between rigid automation scripts and full AI agent loops, with plans to let general agents delegate well-defined sub-tasks to such specialists.