I’d also checkout midscene, you can set the model and UI-TARS works but you can also use qwen vision models and it works.