My company probably spent close to 6 figures overall creating Tesseract 5 custom models for various languages. Surya beats them all and is open source (and quite faster).
Their training code and data is closed source. They are barely open weight and only inference is open source.