Introducing ParseBench-2: The New Standard for Document Parsing->

Introducing OCRia-1.5

A few weeks ago we released OCRia-1, our 4B document parsing model, built with a focus on complex tables and charts.

We didn’t stop there. Today we’re introducing OCRia-1.5. It is more accurate than OCRia-1 on complex tables, charts and forms, and it places bounding boxes more precisely.

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 a new, smaller model that runs the Cost Effective tier. It has half the parameters of OCRia-1, and Cost Effective is now more accurate than before on all six dimensions of ParseBench-2, our document parsing benchmark.

Both models are live in LlamaParse today. With OCRia-1.5 underneath, LlamaParse sets the state of the art on ParseBench-2 for tables, charts and formatting.

More accurate tables, charts, and forms

Complex tables were a main focus of OCRia-1.5. In a financial statement, a header such as “Three Months Ended” can sit above several columns at once. If a parser drops it, the numbers below are still correct, but you can no longer tell whether they cover three months or twelve.

With OCRia-1.5, table scores on ParseBench-2 rose from 86.5 to 90.1 in Agentic and from 83.2 to 87.5 in Cost Effective. The gains were larger on tables with spanning headers like the one above.

Charts and forms improved in both tiers as well. 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, rose by 2.6 points in Agentic and 4.2 points in Cost Effective.

More precise visual grounding

Applications use LlamaParse’s bounding boxes to highlight the source behind an answer. A misplaced box can point to neighboring text or leave part of a table outside the highlight, and the reader can no longer check the answer against the page.

With OCRia-1.5, the boxes follow the structure of the document more closely. Each field on a form gets its own box. When a page has two tables, both are boxed, and each box covers the whole table.

With OCRia-1.5, Cost Effective’s visual grounding score on ParseBench-2 rose from 77.0 to 88.5, and Agentic’s rose by 4.1 points.

Available in LlamaParse

Both OCRia-1.5 models are live in LlamaParse today. The new OCRia-1.5-2B runs Cost Effective at 3 credits per page and suits high-volume parsing. The upgraded OCRia-1.5-4B runs Agentic at 10 credits per page and remains the more accurate choice for complex tables and charts.

Try OCRia-1.5 on your documents.

Related articles

Keep Reading

Start building your first document agent today

LlamaIndex gets you from raw data to real automation — fast.