The company says it is now used in LlamaParse and is intended to recover complex tables, chart data, formatting, and mathematical notation. LlamaIndex says the LlamaParse price remains unchanged.
In the company’s ParseBench comparison, OCRia-1 scored 87.4 overall, 12.1 points ahead of the next system listed. LlamaIndex reports table scores of 91.5 on the ParseBench tables dimension, 92.9 on olmOCR-bench, and 84.1 on PulseBench-Tab. It reports an 85.1 ParseBench chart-extraction score, versus 66.1 for the next model listed. The model can output chart data as HTML tables and preserve formatting in Markdown and HTML, while rendering formulas in LaTeX.
LlamaIndex says its training set contained more than 100,000 document pages, with human annotations, generated and reviewed annotations, synthetic documents, supervised fine-tuning, and reinforcement learning with verifiable rewards. The source does not provide independently reproduced benchmark results, numerical pricing or latency figures, or information on whether model weights are available.