Frontier vision models can read a page of text. What they keep failing at is the part a finance team actually cares about: the nested header spanning four columns, the merged cell in the middle, the stacked bar chart whose series are never labeled. LlamaIndex’s OCRia-1, a 4-billion-parameter vision-language model now powering LlamaParse, is a bet that this gap is a training-data problem rather than a parameter-count problem.
A 12-Point Lead at 4B
On ParseBench, OCRia-1 scores 87.4 overall — text content, formatting, tables and charts — the highest score for any parsing model under 35B parameters, and 12.1 points ahead of the runner-up, LightOnOCR-3-4B at 75.3. Mistral OCR 4 manages 67.5, olmOCR-2 (7B) 43.3.
The lead holds on three independent third-party benchmarks: olmOCR-bench (83.5), PulseBench-Tab (84.1) and fr-bench-pdf2md (72.1). It also leads every table-specific evaluation across all four benchmarks, including the table dimension of ParseBench at 91.5 against Chandra-OCR-2’s 89.2, and the tables category of olmOCR-bench at 92.9. Long tables in fr-bench-pdf2md — 1,820 multilingual tables in PulseBench-Tab — are where structure usually collapses, and OCRia-1 holds 83.2 and 81.9 in both long-table categories.
Charts Are a Data Problem, Not a Summary Problem
The chart number is the most interesting. OCRia-1 scores 85.1 on ParseBench’s chart dimension: a 19-point lead over LightOnOCR-3-4B at 66.1, while conventional OCR approaches land near zero — Mistral OCR 4 scores 40.1, Surya OCR 2 scores 1.2, dots.mocr 0.1.
The reason is the output format. A chart gets converted into an HTML table of its data points, with each value bound to the right series and label. “Revenue grew across all segments” is useless for analysis; a table of the underlying numbers is not. Scoring requires a value to appear under the correct row and column headers, so a parser cannot pass by reading the axes plausibly.
Formatting is the third dimension: 83.4, against 67.9 for the next system. Superscripts, subscripts, strikethroughs and form checkboxes survive into Markdown and HTML, and formulas come out as LaTeX instead of flattened text.
Where the Model Came From
Training data is the quiet story. The set is 100k+ document pages aimed at what modern VLMs still get wrong — dense multi-page tables, charts, mathematical notation, multilingual text, difficult scans. Ground truth was bootstrapped with generation-and-review VLMs plus human annotation, and synthetic document generators fill in the long tail of real-world complexity.
Post-training combines supervised fine-tuning for baseline parsing with RLVR — reinforcement learning against verifiable rewards — targeted at generalization on complex tables and charts. Grading a parsed table against ground truth is checkable by machine, which is exactly what RLVR needs to exist at all.
The result is a different shape of win than the usual model launch. Not a bigger model that does everything a bit better, but a small specialist that does one thing measurably better than everything else, at a size that can be served in production.