OCR for redaction tasks are more difficult for VLMs in that accurate bounding boxes for every word on a page are essential to correctly obscure words on a page. Until recently, most VLMs (particularly open source) have not been good at this task. Early in February, I posted here my tests with Qwen 3
