LlamaIndex, the company behind the LlamaIndex framework and the LlamaParse document service, introduced OCRia-1 on October 22, a vision-language model for converting document pages into structured text. With 4 billion parameters, it is small by current standards, and LlamaIndex is positioning it as the most accurate parser it has tested on tables, charts and formatting, while also being fast enough to process very large archives on a single GPU.
The headline number comes from ParseBench, a benchmark of enterprise pages from insurance, finance and government documents. OCRia-1 scores 87.4 overall there, against 75.3 for LightOnOCR-3-4B, the next system in LlamaIndex's comparison, and 67.5 for Mistral OCR 4.1. The gaps are wider on specific skills: 85.1 on chart extraction versus 66.1 for the runner-up, and 83.4 on preserving formatting such as superscripts, strikethroughs and checkboxes, versus 67.9. ParseBench was built by LlamaIndex researchers, so these are results on the company's own yardstick.
LlamaIndex also reports first place on table evaluations from three outside benchmarks: 92.9 on the tables category of olmOCR-bench, 84.1 on the 1,820 multilingual tables of PulseBench-Tab, and the two long-table categories of fr-bench-pdf2md. Some margins there are narrow. On the overall olmOCR-bench score, OCRia-1 reaches 83.5 against 82.3 for the 7B olmOCR-2, and on fr-bench-pdf2md it leads LightOnOCR-3-4B by less than a point.
Technically, the model is a fine-tune of Qwen3.5-4B. It writes pages as Markdown, renders tables as HTML with merged cells and nested headers intact, turns charts into tables of their underlying data points and transcribes formulas as LaTeX. According to the technical report, most of the effort went into data: a training set of more than 100,000 pages mixing real documents, checked against the PDF text layer, with generated pages whose correct output is known in advance, which helps with lengthy tables and heavy formatting. Training combined supervised fine-tuning with reinforcement learning using verifiable rewards scored separately for tables, charts and formatting.
Speed is the other half of the pitch. LlamaIndex says the served model decodes about 4,700 tokens per second for a single request on one NVIDIA B200, roughly nine times the base model on stock vLLM, thanks to a block-diffusion speculative drafter, mixed NVFP4 and FP8 quantization and custom kernels. At 64 concurrent requests it parses about ten pages per second, which the report translates into roughly 860,000 pages a day per GPU.
For users, the practical change is inside LlamaParse, where OCRia-1 now runs at the same price as before. That matters most for retrieval pipelines and document agents in finance and insurance, where a number placed under the wrong column header produces a confident but incorrect answer downstream.
There are limits to what can be checked today. The model is available only through LlamaIndex's hosted service; the report says a version will be made available soon, without a date or licence. All comparisons were run by LlamaIndex, and the strongest leads are on a benchmark it created, so independent evaluations will be needed to confirm how far ahead OCRia-1 really is.