LlamaIndex on November 18, 2026 introduced OCRia-1.5, a new generation of the document parsing models that power its LlamaParse service. The post, written by Gauthami Polasani, says the update is more accurate on complex tables, charts and forms and places bounding boxes more precisely than OCRia-1, the 4B-parameter model the company released a few weeks earlier with a focus on complex tables and charts.
OCRia-1.5 comes in two sizes. OCRia-1.5-4B is the upgraded flagship and runs LlamaParse's Agentic tier. OCRia-1.5-2B is new, has half the parameters of OCRia-1 and runs the Cost Effective tier. Both models are live in LlamaParse today. The 2B model costs 3 credits per page and is aimed at high-volume parsing, while the 4B model costs 10 credits per page and, according to LlamaIndex, remains the more accurate choice for complex tables and charts.
The company reports its gains on ParseBench-2, its own document parsing benchmark. Table scores rose from 86.5 to 90.1 in the Agentic tier and from 83.2 to 87.5 in Cost Effective, with larger gains on tables that have spanning headers. LlamaIndex gives the example of a financial statement in which a header such as "Three Months Ended" sits above several columns: if a parser drops it, the numbers are still right, but readers can no longer tell whether they cover three months or twelve. Chart scores rose from 74.9 to 76.8 in Agentic and from 67.3 to 73.0 in Cost Effective. The forms score, which checks that values stay attached to the right fields and that checkboxes keep their state, improved by 2.6 points in Agentic and 4.2 points in Cost Effective.
The second focus is visual grounding. Applications use LlamaParse bounding boxes to highlight the source behind an answer, and a misplaced box can point at neighboring text or leave part of a table outside the highlight, which makes the answer hard to check against the page. With OCRia-1.5, LlamaIndex says each form field gets its own box, and when a page holds two tables, both are boxed and each box covers the whole table. On ParseBench-2, the Cost Effective grounding score rose from 77.0 to 88.5 and the Agentic score by 4.1 points. LlamaIndex adds that Cost Effective is now more accurate than before on all six benchmark dimensions despite the smaller model.
On the strength of these results, the company says LlamaParse with OCRia-1.5 underneath sets the state of the art on ParseBench-2 for tables, charts and formatting.
For teams building retrieval and agent pipelines over financial reports, invoices and scanned forms, the practical relevance is in two places. Better table structure reduces silent errors in extracted numbers, and tighter boxes make it easier to show users exactly where an answer came from. The cheaper 2B tier also gives a lower-cost option for large batches.
There are limits to what can be concluded. The benchmark is LlamaIndex's own and the comparisons are against its previous model, so the claim of state-of-the-art performance has not been independently tested. The post does not say whether the model weights are released or under what license, and it gives no results for other vendors' systems. The source is short, so this article does not go beyond the details the company published.