It targets complex tables, charts, scanned documents and formatting preservation.
Technical specifications
| Metric | OCRia-1 |
| Parameters | 4B |
| ParseBench | 87.4 |
| olmOCR-bench | 83.5 |
| olmOCR-bench Tables | 92.9 |
| PulseBench-Tab | 84.1 |
| fr-bench-pdf2md | 72.1 |
| Training data | 100,000+ document pages |
| Output | Markdown, HTML, LaTeX |
| Deployment | LlamaParse |
Training used supervised fine-tuning and RLVR, with emphasis on difficult layouts, scanned pages, tables and mathematical notation.
Table parsing
OCRia-1 is optimized for:
- nested headers and merged cells;
- irregular and scanned tables;
- long tables;
- unlabeled charts and technical diagrams.
Its strongest result is 92.9 on the table subset of olmOCR-bench.
Comparison with similar 4B models
There is no publicly documented OCRia predecessor, so the closest comparison is with other models in the same parameter class.
| Model | ParseBench | olmOCR-bench | PulseBench-Tab |
| OCRia-1 | 87.4 | 83.5 | 84.1 |
| LightOnOCR-3 | 75.3 | 81.2 | 82.4 |
| Chandra-OCR-2 | 74.8 | 81.2 | 80.0 |
On ParseBench, OCRia-1 leads LightOnOCR-3 by 12.1 points.
For complex tables, the gap is also significant: OCRia-1 scores 91.5 on ParseBench Tables versus 82.5 for LightOnOCR-3.
OCRia-1's main advantage is structural parsing, not just text recognition.
Formatting and charts
The model preserves:
- headings and document hierarchy;
- superscripts, subscripts and checkboxes;
- mathematical expressions;
- Markdown and HTML structure.
It also parses line, bar and pie charts as well as technical diagrams.
Availability
OCRia-1 now powers LlamaParse's Cost-effective, Agentic and Agentic Plus modes. The company positions the model for high-volume enterprise workloads, including processing millions or billions of pages.
