olmOCR 2 arrives in the Ai2 Playground with open weights

Ai2's July 2026 announcement added olmOCR 2 to the Ai2 Playground, its web interface for trying its open text, image and video models. The model itself was released in October 2025 with its weights, training data and code, so the Playground adds a way to test it without running it yourself.

olmOCR 2 converts an image of a document page directly into structured text: headings as Markdown, tables as HTML and mathematical notation as LaTeX. That matters because ordinary character recognition can get every word right and still scramble a two-column layout or flatten a table, which changes what the text means to a search system or an analysis tool. The model is a fine-tuned version of Qwen2.5-VL-7B, an open model that reads images and text. Its distinctive training step scores each output with automatic pass-or-fail tests that check properties such as whether table structure is preserved, math is faithfully transcribed and reading order is consistent, and rewards outputs that pass more of them. Many of those tests come from synthetic pages with known correct answers, made by having Claude Sonnet 4 rebuild real pages as HTML.

On olmOCR-Bench, Ai2's own test set, olmOCR 2 scored 82.4, nearly four points above the previous olmOCR release, with the largest gains on tables, multi-column pages and old scanned math. The paper describes that benchmark as English, so the score says little about documents in other languages.