# Datalab releases Surya 2.5 OCR model with 85.9% accuracy on OmniParseBench

Digest AI · Generative AI & Models · published 2026-10-30T17:55:46Z

Canonical: https://digestai.news/story/datalab-releases-surya-2-5-ocr-model-with-85-9-accuracy-on-omniparsebe

## Summary

Datalab released Surya 2.5, a 689M-parameter open-source OCR model that outperforms its predecessor on key benchmarks. It scores **85.9%** on OmniParseBench and **83.7%** on olmOCR-bench, surpassing Surya 2’s **83.3%** and **87.2%** respectively. The model also achieves **74.9%** multilingual accuracy across 91 languages, with 21 scoring ≥90% and 52 ≥80%. Surya 2.5 combines layout analysis, OCR, and table recognition into a single vision-language model (VLM), improving throughput to **4.5 pages/second** on an H100 GPU at 192 DPI—**1.7x faster** than Surya 2 with MTP enabled. It outputs structured HTML for tables and reading order, supports math inline via LaTeX, and includes smaller models for line-level detection and error correction. The project provides a managed platform with $5 free credits and a Streamlit app for interactive testing. The model weights use a modified AI Pubs Open Rail-M license, with commercial licensing available separately. Since its launch, Surya has seen **1,222 downloads last month** on Hugging Face.

## Key points

- Surya 2.5 scores **85.9%** on OmniParseBench, **83.7%** on olmOCR-bench, and **74.9%** multilingual accuracy across 91 languages
- Throughput improved to **4.5 pages/second** on an H100 GPU (192 DPI), **1.7x faster** than Surya 2 with MTP
- Outputs structured HTML for tables, math (LaTeX), and reading order; includes smaller models for line detection and error correction

## Why it matters

Surya 2.5’s accuracy and speed make it a strong open alternative for document parsing, benefiting research, archival digitization, and multilingual workflows. Its unified VLM approach reduces inference overhead, while structured outputs simplify downstream processing for tables, math, and layout.

## Sources

1. [datalab-to/surya-ocr-2.5 · Hugging Face](https://huggingface.co/datalab-to/surya-ocr-2.5) (huggingface.co, 2026-10-30, primary source)

## Cite

Digest AI, "Datalab releases Surya 2.5 OCR model with 85.9% accuracy on OmniParseBench", 30 October 2026, https://digestai.news/story/datalab-releases-surya-2-5-ocr-model-with-85-9-accuracy-on-omniparsebe

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/datalab-releases-surya-2-5-ocr-model-with-85-9-accuracy-on-omniparsebe.json