GitHub radar
Read scanned documents and tables locally, no cloud
A small model for parsing scanned and photographed documents pulls out text and tables more accurately than several much larger general models, while running on a regular laptop.
TeleOCR is a compact model for document parsing. It comes from the team XingChen-AGI, the model was earlier named NaviDC-OCR. It is trained to read more than clean digital scans. It also handles documents photographed at an angle or with distortion, without a separate page-straightening step. TeleOCR recognizes not just text but also structure, tables, formulas, and the reading order of a multi-column page. At around 1.2 to 1.4 billion parameters, it scored 96.87 on the OmniDocBench v1.6 benchmark. That is higher than much larger general-purpose models, including Gemini 3 Pro and GPT-5.2. The model is built for this one task. It does not hold a conversation and does not write code, it only takes in a document and returns marked-up text.
Why a vibe-coder should care
A vibe coder sometimes needs to pull text and a table out of a scanned contract, a receipt, or a photographed page. TeleOCR is a ready local tool for exactly that, instead of a reason to build a separate service or pay for a cloud OCR API. The model is small, so it can stay installed and run over batches of documents. It is not suited for chatting with an agent or writing code, its only job is pulling text and structure out of a document.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the document-parsing model TeleOCR from this page: https://huggingface.co/XingChen-AGI/TeleOCR — figure out how to run it on my machine and show me how to feed it a scanned document or photo and get back text and tables.
A regular laptop with 8 GB of RAM or more — the model itself takes up about 1 GB.
Open on Hugging Face▌ More finds