LlamaIndex Launches OCRia-1, a 4B Document-Parsing Model Claiming Top ParseBench Score
The 4B vision-language model claims 87.4 on ParseBench and up to about 860,000 pages per day on one B200 GPU.
LlamaIndex announced OCRia-1, a 4-billion-parameter vision-language model fine-tuned from Qwen3.5-4B that converts document pages into structured text. It outputs Markdown for prose, HTML for complex tables while preserving merged cells and nested headers, tables extracted from charts, and LaTeX for formulas. The company claims an overall score of 87.4 on ParseBench, 12.1 points above the next-best parser, LightOnOCR-3-4B at 75.3 and Mistral OCR 4.1 at 67.5.
According to LlamaIndex, the served model reaches roughly 4,700 tokens per second per request on a single NVIDIA B200 GPU using block-diffusion speculative drafting and NVFP4/FP8 quantisation, equal to about 10 pages per second at 64 concurrent requests. ParseBench covers about 2,000 human-verified enterprise pages from insurance, finance and government. The model is already integrated into LlamaParse across all tiers with no change in per-page pricing. Benchmark figures are vendor-reported.