Tiny specialized models are really interesting.
Aside from agentic coding, this seems to be the most interesting niche to me. There are now lots of models for doing stuff like tts, stt, ocr, and it’s really handy for sifting through documents, extracting transcripts, etc. And it’s all stuff you can run locally on moderate hardware.
o7 Farewell, Tesseract
What languages does it support? Any source?
I don’t see language being passed in the example code, so it may be agnostic?
Isn’t that almost impossible? It has to be trained on that language afaik
I use the inbuild OCR in ShareX, which works pretty well, without AI. FOSS
I find traditional OCR apps are just qualitatively worse. This tool understands complex tables, extracts mathematical formulas, predicts reading order, and outputs nicely formatted Markdown. So, it’s not just blindly scraping flat text from a document.
Why do you think that it doesn’t use AI? It uses Wndows’ own OCR APIs, requiring Windows language packs for support for different scripts & languages. OCR has used AI for a long time now.
Yes, like my chess computer in the 90th. I don’t have any lenguage modell in Windows, nor any other MS crap and services, the first I deleted from Windows. The lenguage pack and fonts are another thing, nothing to do with LLM or AI
