LlamaIndex's OCRia-1 Parses PDFs 75x Faster Than Gemini at a Third the Cost

LlamaIndex has released OCRia-1, a 4-billion-parameter vision-language model that converts document pages into structured Markdown and HTML. The system combines optical character recognition with visual layout analysis, allowing it to preserve tables, charts, formulas, formatting, and reading order.

OCRia-1 now powers LlamaParse and is available through OpenDocRouter, where developers can compare it with Claude Opus, Gemini 4, GPT-6, Mistral OCR, and other document parsers on the same files. Its reported advantage centers on documents where structural mistakes can corrupt downstream retrieval, extraction, and analysis.

A 4B model leads the reported field

In LlamaIndex’s published ParseBench results, OCRia-1 scores 87.4 across text content, formatting, tables, and charts. That is the highest reported score among models under 35 billion parameters and 12.1 points above the nearest model in the comparison.

Beyond ParseBench, LlamaIndex reports leading results on three external evaluations, each with its own document corpus and scoring method:

• olmOCR-bench: 92.9
• PulseBench-Tab: 84.1
• fr-bench-pdf2md: highest score in the tables category

Comparisons belong within each benchmark because the test documents and scoring rubrics differ. Production accuracy will also depend on scan quality, language, page rendering, layout, and the types of documents in a given workload.

Tables keep their meaning

OCRia-1 reconstructs nested headers, merged cells, and irregular grids while preserving the relationships between values and labels. Those relationships matter in financial statements, insurance filings, and scientific tables, where assigning a value to the wrong row or column can produce a plausible but incorrect result.

For charts, the model emits HTML tables containing the underlying data points and their associated series and labels. ParseBench awards credit only when a value appears under the correct row and column headers. OCRia-1 scores 85.1 on that evaluation, 19 points above the next model’s 66.1, while the reported conventional-parser baselines score near zero.

Inline document details also survive the conversion. OCRia-1 retains superscripts, subscripts, strikethroughs, and form checkboxes in Markdown or HTML, and it renders mathematical formulas as LaTeX.

Training targets structural errors

LlamaIndex says the training corpus contains more than 100,000 document pages selected to expose common parsing failures, including dense multipage tables, charts, mathematical notation, multilingual text, and degraded scans. Its training process combines several techniques:

• Targeted examples: Pages emphasize layouts and visual details that current vision-language models frequently misread.
• Reviewed annotations: Ground truth combines human labeling with VLM-based generation and review.
• Synthetic documents: Generators create uncommon layouts and edge cases that appear infrequently in collected data.
• Supervised fine-tuning: SFT teaches the base parsing behavior from labeled input and output pairs.
• Verifiable rewards: Reinforcement learning with verifiable rewards, or RLVR, improves table and chart extraction through automated checks.

RLVR assigns rewards by comparing model output with a target that software can verify. Document parsing supports this approach because a program can check whether a cell appears under the correct header, whether a chart value retains its series label, and whether required structure is present.

OpenDocRouter reports 0.2 seconds per page

OpenDocRouter’s ParseBench run gives OCRia-1 an overall score of 84.67, close to meta/muse-wonder-1.0-minimal at 84.73 and anthropic/claude-opus-5-5 at 84.27. The router’s 84.67 result differs from the 87.4 published by LlamaIndex, so each figure should be interpreted within its own run. Benchmark revisions, parser settings, rendering choices, and execution harnesses can affect the result.

Among these entries, OCRia-1 has the lowest listed cost and page time. Its 0.2-second result is 42 times faster than Muse, 51 times faster than Claude Opus, and about 75 times faster than Gemini. The listed cost equals 33% of Muse’s, 27% of Claude Opus’s, and 68% of Gemini’s.

OpenDocRouter lists token prices of $0.97 per million input tokens and $4.83 per million output tokens, with a quoted worst-case charge of about $0.18 per page. At the leaderboard average of $13.04 per 1,000 pages, a $5 account credit covers roughly 380 pages. Actual usage depends on page complexity and output length.

Two hosted ways to run it

1. LlamaParse now uses OCRia-1 as its parsing engine, with no reported change to the service’s price.
2. OpenDocRouter accepts the model identifier in a parse request:
   code

   {
     "model": "llamaindex/ocria-1"
   }

Test structure and downstream accuracy

Production testing should use a representative corpus that includes clean PDFs, scans, multipage tables, charts, formulas, rotated pages, multilingual text, and the rare layouts that generate support incidents. A useful evaluation should measure:

• Text accuracy and reading order
• Cell-to-header and chart-value associations
• Formula, checkbox, and inline-format preservation
• Valid Markdown, HTML, and LaTeX output rates
• Median and tail latency
• Token use, page cost, retries, and timeouts
• Accuracy in downstream retrieval and extraction tasks
• Manual correction time for failed pages

Pipelines that process financial statements, insurance filings, scientific papers, or chart-heavy reports align closely with OCRia-1’s reported strengths. A corpus-level comparison against the current parser will show whether its benchmark lead translates into fewer structural errors, lower costs, and faster end-to-end processing.