LlamaIndex has released OCRia-1, a 4-billion-parameter vision-language model that parses document pages into structured text for developers and operators running agents over financial reports, insurance filings, tax forms, and other layout-heavy files. It now powers every mode in LlamaParse, the company's parsing service, at the same per-page price as before. Demand for this kind of parsing is rising as more professional work moves into document agents, and most of the information in those files lives in tables, charts, and formatted text that still trip up both large general models and existing specialized parsers. On ParseBench, LlamaIndex's benchmark of enterprise pages, the company reports an overall score of 87.4, 12.1 points above the next system in its comparison. Chart extraction scores 85.1 and formatting preservation scores 83.4 on that same test. LlamaIndex also reports first place on three third-party benchmarks, with a wider gap on tables than on overall scores: 92.9 on olmOCR-bench tables (83.5 overall), 84.1 on PulseBench-Tab, and 72.1 on fr-bench-pdf2md. OCRia-1 reconstructs nested headers, merged cells, and irregular grids, converts unlabeled charts into tables of their data points, keeps superscripts, strikethroughs, and checkboxes, and writes formulas as LaTeX. The model is a fine-tune of Qwen3.5-4B. LlamaIndex reports about 4,700 tokens per second on one B200 GPU and about ten pages per second with 64 concurrent requests. Jerry Liu, who announced the release, and Simon Suo co-founded the San Francisco company in 2023 after Liu's open-source library for connecting private data to language models gained a large developer following. OCRia-1 is now the base model behind LlamaParse.