LlamaIndex OCRia-1: A 4B Document Parser That Claims 87.4 on ParseBench

LlamaIndex has introduced OCRia-1, a 4-billion-parameter vision-language model built to parse hard documents: scanned PDFs, nested tables with merged cells, unlabeled charts and math. The company says it sets a new high on ParseBench with 87.4, 12.1 points ahead of the next-best parsing model, and leads table parsing on three third-party benchmarks. The source is LlamaIndex's own announcement, Introducing OCRia-1, and it comes with a caveat that matters: ParseBench is LlamaIndex's benchmark.

This guide covers what OCRia-1 is, how to read each number, what the training recipe was, and what to test before using it in a document pipeline.

TL;DR

What the model is designed to do

LlamaIndex frames the problem as one of agentic automation: most information in financial reports, insurance filings and tax forms is organized as tables, charts, diagrams and formatted text, and both frontier and open-weight models "still struggle to parse them accurately, cheaply, and quickly at scale." OCRia-1 targets three dimensions: accuracy, cost and latency.

Specific capabilities named in the post:

• Complex tables. It reconstructs nested headers, merged cells and irregular grids, preserving the relationships between values and labels. Losing those links can turn an accurate transcription into a wrong financial comparison.
• Charts to data. It converts charts, including stacked bars, overlapping lines and multiple axes, into HTML tables of data points, keeping values attached to the right series and labels.
• Formatting and math. It retains superscripts, subscripts, strikethroughs and form checkboxes in Markdown and HTML, and transcribes formulas as LaTeX.

A green lens over a page divided into rectangular regions, illustrating document layout parsing with OCRia-1

The benchmark results in detail

LlamaIndex's overall results across four benchmarks (higher is better):

Notice the shape of the lead. On ParseBench the margin is 12 points, but on the three other benchmarks it is 0.7 to 1.2 points overall. The big ParseBench gap reflects dimensions that ParseBench emphasizes, charts and formatting, where the post reports gaps of 19.0 and 15.5 points. On tables, the lead is smaller and more consistent: 91.5 vs 89.2 on ParseBench tables, 92.9 vs 92.1 on olmOCR-bench tables, and 83.2 and 81.9 on two long-table categories of fr-bench-pdf2md versus 80.6 and 77.8 for the next best.

Other ParseBench rows in the chart: Mistral OCR 4.1 at 67.5, LightOnOCR-2-1B at 61.1, Surya OCR 2 at 58.4, dots.mocr at 55.6, and olmOCR-2 at 43.3.

How to read the claims

1. ParseBench is LlamaIndex's benchmark. The company released it in April 2026 to evaluate document parsing for agents, with about 2,000 human-verified pages across dimensions such as tables, charts and formatting. A benchmark built by a vendor can still be well designed, but a vendor winning its own benchmark is weaker evidence than a win elsewhere. The third-party results are the better signal, and there the margins are small.
2. Scores for olmOCR-bench were adjusted. The post flags the olmOCR-bench scores as adjusted, so they may differ from what each project publishes.
3. Check the comparison set. The models we saw listed are specialized and open parsers plus Mistral OCR; we did not see frontier general-purpose VLMs in the figures we could read, so confirm how OCRia-1 compares with the largest models on the same pages.
4. No latency or cost table in the post. The announcement says the model is best on accuracy, cost and latency, and that it arrives in LlamaParse at the same price. The post provides no speed or price figures to verify.
5. Independent testing is needed. Test on your own documents: your worst scans, your table formats, your languages.

Background: what ParseBench measures

LlamaIndex released ParseBench in April 2026 as, in its words, the first document parsing benchmark for AI agents; its announcement and the arXiv paper describe about 2,000 human-verified enterprise document pages with more than 167,000 test rules, organized around five dimensions: tables, charts, content faithfulness, semantic formatting and visual grounding. At launch, the paper evaluated 14 methods including vision-language models, specialized parsers and LlamaParse, and reported that general VLMs were competitive on content extraction but weak on chart recovery and visual grounding. A later leaderboard with Kaggle opened the benchmark to outside submissions.

The OCRia-1 post reports four ParseBench-based figures: overall 87.4, tables 91.5, charts 85.1 and formatting 83.4. The charts number is the most striking: the post puts it 19.0 points ahead of the next-closest model (66.1). Chart-to-data extraction is a capability many OCR tools do not attempt, so check how the benchmark scores systems that output no chart data.

Why this class of model matters for agents

Agents that read contracts, invoices, filings or medical forms are only as good as the text and structure they are handed. A wrong cell in a table or a lost footnote becomes a wrong answer downstream, and the error is quiet because the language model sounds confident. That is why document parsing has become its own product category, separate from general chat models, and why a 4B model that runs cheaply per page can be attractive: a pipeline might parse millions of pages. The trade-off for teams is between three options: send pages to a frontier model (flexible, expensive, slow), run a small specialized parser such as OCRia-1 (cheap and fast, narrower), or use classic OCR plus rules (cheapest, brittle on tables and charts).

For retrieval-augmented generation, structure preservation also changes chunking. A parser that outputs clean Markdown or HTML tables lets you chunk by section and keep each table whole, instead of splitting a table across chunks where its header row is lost. If you are building such a pipeline, run the same 100 pages through your current parser and OCRia-1 and diff the table outputs by hand before you commit.

Training recipe

LlamaIndex says it assembled more than 100,000 document pages aimed at cases modern VLMs still get wrong: dense, multi-page tables and charts, math, multilingual text and difficult scans. Ground truth was bootstrapped with generation and review VLMs together with human annotation, and synthetic document generators supplemented real data to vary complexity and match the long tail of real documents. Post-training used supervised fine-tuning for a base layer of parsing ability and reinforcement learning with verifiable rewards (RLVR) to improve generalization on tables and charts. A technical report is linked from the announcement.

Where it sits among OCR models

Document OCR has become a crowded field of small specialized models. On this blog we covered LightOnOCR-3 and ParseBench, Cohere Parse 5, Tencent's Youtu-Parsing Omni, Interfaze-1 Lite and the CPU-friendly Indra OCR. LlamaIndex's earlier work on classification and splitting is in our DocJev write-up. The notable difference with OCRia-1 is that LlamaIndex is a platform company: the model's main distribution is through its own paid parsing service, which is why the question of open weights matters to developers.

How to evaluate it

A quick decision guide

What is still unknown

• Whether the weights are downloadable, and under what license.
• Latency and throughput numbers under stated hardware and settings.
• Performance on handwriting and low-resource languages beyond the benchmarks listed.
• How results change when baselines are run with their recommended rendering rather than a common one.

Related reading

• LightOnOCR-3 and ParseBench
• Cohere Parse 5 on ParseBench
• Tencent Youtu-Parsing Omni
• Indra OCR on CPU
• LlamaIndex DocJev
• Official: Introducing OCRia-1 and LlamaIndex

Details are accurate as of October 22, 2026 and come from LlamaIndex's announcement. Benchmarks are the vendor's own.