LlamaParse

Overview

LlamaParse is LlamaIndex document parsing service. It turns PDFs, scans, Office files, spreadsheets and images into Markdown, structured JSON or page data with coordinates that retrieval systems and document agents can consume directly. On October 22, 2026, LlamaIndex released OCRia-1, a 4B vision language model fine-tuned from Qwen3.5-4B, and it took over LlamaParse parsing the same day.

On ParseBench, the benchmark LlamaIndex maintains itself, OCRia-1 scores 87.4, 12.1 points above the next system LlamaIndex lists, LightOnOCR-3-4B. The gap is widest on charts at 85.1 against 66.1 and on formatting at 83.4 against 67.9. It also leads the table categories of olmOCR-bench, PulseBench-Tab and fr-bench-pdf2md, where overall margins sit around one point. Weighing these numbers means weighing who ran the benchmark.

Billing runs on credits per page across four parsing tiers, from 1 credit to 45 credits per page, at $1.25 per 1,000 credits. The free plan includes 10,000 credits a month and paid plans start at $50 a month with 40,000 credits. OCRia-1 currently runs inside LlamaParse as a hosted service, and its weights have not been published.

Key Features

• Four parsing tiers by complexity: Fast costs 1 credit per page and returns positioned text only. Cost-effective costs 3, Agentic 10 and Agentic Plus 45. Higher tiers lean further on vision models for layout, tables and charts, and switching is a single parameter on the same endpoint.
• Auto Mode routes page by page: Each page is routed to a tier based on its layout complexity, so simple pages stay on the cheap tier and only dense ones step up. LlamaIndex puts the saving at up to 80 percent. You can also pin a tier and parse only the page ranges you need.
• Tables and charts become structured data: Nested headers, merged cells and irregular grids keep the relationships between values and labels, with tables emitted as HTML. Charts turn into tables of their data points, with values bound to the right series and axis labels so they can feed a calculation.
• Layout and semantic formatting preserved: Results carry paragraph level bounding boxes and structure labels that separate headings, tables, lists, formulas and footnotes. Superscripts, subscripts, strikethroughs and form checkboxes survive in Markdown and HTML, and math is transcribed as LaTeX. Layout extraction adds 3 credits per page.
• Many output formats, broad file support: Output can be Markdown, plain text, per-page JSON, XLSX, HTML tables or an annotated PDF, and parsing can target a custom schema for structured JSON. More than 130 file formats are accepted, covering PDF, Office documents, spreadsheet files and common image formats.
• Extraction, classification and splitting alongside parsing: Extract pulls target fields per document, per page or per table row and returns citations with confidence scores. Turbo handles it in one pass at 35 credits per page. Classify routes documents by type at 1 to 2 credits per page and Split breaks bound volumes apart at 4, so filtering first keeps the expensive tiers for what needs them.

Use Cases

• Turn table-heavy scans of financial statements, policies and tax forms into Markdown or JSON that can be computed on, without header drift sending the wrong number downstream.
• Preprocess documents for retrieval augmented question answering while keeping structure labels and bounding boxes, so an answer can point back to a spot on the page.
• Batch extract invoice and contract fields such as numbers, amounts and terms with confidence scores, and send low-confidence items to human review.
• Normalise contracts, official letters and technical manuals in many languages into one structured format for archiving.
• Recover formulas and charts from papers and patents, with math as LaTeX and charts as data tables ready for analysis.
• Classify by document type first and only send the parts that matter into the expensive tiers, keeping the parsing budget under control.

Pros

• Per-page credit pricing with selectable tiers, so testing can start on the 3-credit tier that covers most paperwork.
• Reparsing the same file within 48 hours is free, which keeps pipeline debugging affordable.
• Parse once and extract many times, since extraction on an already parsed file is billed only at the extract tier.
• Output carries paragraph bounding boxes and structure labels, so retrieval hits can be located on the page.
• The free plan includes 10,000 credits monthly, paid plans start at $50 a month with 40,000 credits, and enterprise adds hybrid cloud deployment and higher concurrency.
• Over 130 file formats and more than a hundred languages are covered, so documents from different sources do not need separate tools.

Pricing

Billing is credit based at $1.25 per 1,000 credits. Parsing has four tiers: Fast at 1 credit per page, Cost-effective at 3, Agentic at 10 and Agentic Plus at 45. Layout extraction adds 3 credits per page on any tier, and enriched form output adds 10 credits per page that contains a form. Extraction is billed by tier on top: Agentic Plus 50 credits per page, Agentic 15, Cost-effective 5, while Turbo covers the whole job at 35. Splitting costs 4 credits per page, classification 1 to 2, indexing export 2, retrieval queries 1 each and chat turns 100. Plans: free with 10,000 credits a month, Starter at $50 a month with 40,000 credits, Pro at $500 a month with 400,000 credits and custom enterprise terms. See the official site for current quotas and concurrency limits.

Summary

LlamaParse is document preprocessing for agents and retrieval systems, turning complex layouts into structured data that can be computed on. With OCRia-1 behind the parser, table and chart recovery becomes the main line, and four billing tiers let teams spend by document complexity. It suits teams working through table-heavy financial, insurance and invoice documents.

Version History

• OCRia-1 takes over LlamaParse parsing with 87.4 on LlamaIndex own ParseBench (2026-10-22): LlamaIndex released OCRia-1, a 4B document parsing model fine-tuned from Qwen3.5-4B, on October 22, 2026, and it now runs LlamaParse parsing. On ParseBench, the benchmark LlamaIndex maintains itself, OCRia-1 scores 87.4, 12.1 points above the next system LlamaIndex lists, LightOnOCR-3-4B, with 85.1 against 66.1 on charts and 83.4 against 67.9 on formatting. On outside benchmarks it reaches 83.5 overall on olmOCR-bench, 84.1 on PulseBench-Tab and 72.1 on fr-bench-pdf2md. A single B200 decodes about 4,700 tokens per second on one request and parses roughly ten pages per second at 64 concurrent requests, which LlamaIndex puts at about 860,000 pages per GPU per day. LlamaParse pricing tiers are unchanged and the model weights have not been published.