4B Model Achieves Document-Parsing SOTA; LlamaIndex Releases 37-Page Technical Report

Document parsing is the least sexy but most critical step in putting LLMs into production. It turns tables, charts, and text in PDFs and scans into structured data, and it is the prerequisite for Agents to process real business documents. Today LlamaIndex released OCRia-1: a 4B-parameter vision-language model that scores 87.4 on ParseBench, 12.1 points ahead of the second-place model. It is not 70B, not a multimodal behemoth—just 4B. At the same time, it also took first place on three other third-party benchmarks: olmOCR-bench overall 83.5 and table category 92.9; PulseBench-Tab 84.1; fr-bench-pdf2md 72.1.

The model focuses optimization on the areas that hurt most in enterprise documents: complex tables and charts. Nested headers, merged cells, irregular grids, and scanned tables—OCRia-1 can reconstruct the structure. It scores 91.5 on the table dimension of ParseBench. For people processing financial reports, tax forms, and statements, this capability directly affects whether downstream calculations are correct. It is also good at converting unlabeled charts into HTML tables with correct values and correspondences, scoring 85.1 on the chart-extraction dimension, 19 points higher than the second place. Format preservation is not neglected either: superscripts, subscripts, strikethrough, checkboxes, and LaTeX transcription of math formulas all contribute to an 83.4 score on the ParseBench formatting dimension.

On speed, the team claims decoding is more than 2–4x faster than comparable open-source models of the same class. A 4B parameter count means lower inference cost; at large-scale document processing, this gap will show up directly on the bill. The technical report is 37 pages, 24 of which are appendices documenting data filtering, training experiments, and evaluation details. For people who want to train their own vision models, these appendices may be more valuable than the scores.

OCRia-1 is already live in LlamaParse, and all modes use it as the base model. One developer commented: achieving this accuracy with a 4B model is no joke—it is a huge unlock for turning messy documents into structured context. If you are processing millions of documents, you can directly use the official evaluation path.

Document parsing does not get as much spotlight as large models, but it is the first hurdle Agents must clear in production. OCRia-1 reaching this level with 4B parameters shows that this path is not at its end.

More data and technical details are in the technical report and official blog.