Most of the world's high-value text — scientific papers, government filings, scanned books — is locked in PDFs that classic OCR mangles into garbled, out-of-order fragments. The reframing here is to treat OCR as a vision-language task: a fine-tuned 7B VLM reads each page the way a person would, then emits clean Markdown in natural reading order. The payoff is as much economic as technical — at roughly $200 per million pages, parsing whole document corpora for LLM training becomes financially realistic, about 1/32 the cost of pushing the same pages through GPT-4o.
What Sets It Apart
- Document-anchoring prompt: it pairs the rendered page image with text and metadata already inside the PDF, so the model grounds its output instead of hallucinating — meaning fewer dropped equations and mangled tables on messy scans.
- Reading-order aware: multi-column layouts, figures and insets come out in the order a human reads them, and repeating headers/footers are stripped — so the text is usable as training data without heavy post-cleaning.
- Built to scale: the toolkit drives SGLang/vLLM from a single GPU up to hundreds, so the same code runs a laptop experiment or a corpus-wide job.
- Open end to end: Apache-2.0 weights, toolkit, and the olmOCR-Bench eval are all public; v0.4 scores 82.4 on that benchmark, competitive with closed pipelines.
Who It's For
Great fit if you are building LLM datasets, RAG corpora, or research archives and need millions of pages parsed locally without per-call API fees. Look elsewhere if you have no GPU (it needs ~12GB+ VRAM), only a handful of clean digital-born PDFs where a plain text extractor is faster and simpler, or you want a turnkey hosted API rather than a self-run toolkit.
