Datalab releases OmniParseBench OCR benchmark

OmniParseBench is an open benchmark that tests optical character recognition and document parsing systems using 16,288 yes-or-no questions across thousands of document pages in 92 languages. The benchmark addresses limitations in existing OCR benchmarks by providing clear, auditable pass-or-fail scoring rather than complex metrics that cannot be easily compared, and by allowing results to be filtered by content characteristics like language, structure, and rendering to match specific document types. The evaluation tool runs offline without requiring API keys and does not use models to judge results, making it deployable on any machine with Python 3.11 or later. Results show that top-performing systems are separated by relatively small margins, with Gemini 3.8 Flash leading at 92.9 and Datalab's own parser following at 92.7.