Try various downloadable weights that has Vision, they're all good at different examples, running multiple ones and then finally something to aggregate/figure out the right one usually does the trick. Some recent ones to keep in the list: ministral-3-14b-reasoning, qwen3-vl-30b, magistral-small-2509, gemma-3-27b

Personally I found magistral-small-2509 to be overall most accurate, but it completely fails on some samples, while qwen3-vl-30b doesn't struggle at all with those same samples. So seems training data is really uneven depending on what exactly you're trying to OCR.

And the trade-off of course is that these are LLMs so not exactly lightweight nor fast on consumer hardware, but at least with the approach of using multiple you greatly increase the accuracy.