olmOCR 7B

olmOCR is AllenAI's open toolkit for converting digitized documents — PDFs, scans, even handwritten pages — into clean, naturally ordered plain text. The core is olmOCR-7B, a 7-billion-parameter vision-language model fine-tuned from Qwen2-VL-7B-Instruct on the olmOCR-mix dataset of 260,000 diverse PDF pages, with a follow-up olmOCR 2 trained by reinforcement learning on synthetic documents with verifiable unit-test rewards.

The headline numbers are striking: on the olmOCR-Bench suite of 1,400 challenging PDFs, the 7B model beats GPT-4o, Gemini Flash 2, and Qwen 2.5-VL on OCR quality — particularly math formulas, tables, and multi-column layouts — while processing a million PDF pages for a reported $176 in compute. Everything is open: model weights, training data, code, and benchmark, all under Apache 2.0.

Honest caveats: it is an OCR specialist, not a general document-understanding agent, and the benchmark is AllenAI's own. The newest olmOCR 2 weights (late 2025) build on the same pipeline. For bulk PDF-to-text pipelines where GPT-4o-class APIs are overkill and overpriced, this is the open default.