LlamaIndex Ranks OCR Models: OCRia-1 Leads Table Benchmarks with 91.5 Score
·2026.11.18 09:00
Key point
LlamaIndex's ParseBench-2 evaluation shows OCRia-1 leads in table and chart processing benchmarks, offering lower cost and faster latency than frontier VLMs like Claude Opus 5.5, which achieve higher raw table accuracy.
Details
- OCRia-1 (LlamaParse's 4B parameter model) ranks first in table processing across four benchmarks, scoring 91.5 on ParseBench tables in the launch run and 90.90 on the public ParseBench leaderboard, with a 2.3-point lead over Chandra OCR 2 in the launch run.
- Frontier VLMs like Claude Opus 5.5 (94.26 table score) and GPT-6.1 Terra (93.07) achieve higher raw table accuracy than OCRia-1 but suffer from high costs (6.13 cents per page for Opus 5.5) and slower latency (11.2s vs. OCRia-1's 2.2s).
- OCRia-1 leads chart processing with an 85.1 ParseBench score, significantly ahead of the next model at 66.1, and returns data points as HTML tables for numerical analysis.
- Open-weight models like Chandra OCR 2 (89.20 table score) and LightOnOCR-3 (83.63) offer self-hosting options, while commercial APIs like AWS Textract and Google Document AI provide lower costs for general text but lower table scores (84.46 and 54.95 respectively).
- ParseBench-2 evaluates models on 2,000 human-verified pages and 169,000+ test rules across five dimensions, cross-validated with external benchmarks like olmOCR-bench and PulseBench-Tab to ensure neutrality.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.