Datalab releases Surya 2.5 OCR model with 85.9% accuracy on OmniParseBench
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…
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
datalab-to/surya-ocr-2.5 · Hugging Face
huggingface.co · 30 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Sensenova/SenseNova-Loop-9B-Preview · Hugging Face · 1 src
- TBC launches Mischief video model with neuron-inspired speed boost · 1 src
- NVIDIA releases Metatron v1 unified humanoid control policy for Unitree G1 · 1 src
- Edge0 releases Audio8-TTS-Realistic, a 0.6B open-weight voice cloning model · 1 src
- Google launches Gemini 4 Carbon for cybersecurity defenders via Fairwind at $2/$10 per million tokens · 12 src
Comments
via GitHub Discussions