For OCR applications in offline environments, edge devices, or data-privacy-sensitive scenarios, a 67M model that runs on CPU may be more suitable than calling a cloud-based large model API.