Surya 2.5

Surya 2.5 is Datalab’s 689M-parameter open OCR / layout model released October 29, 2026. It reads a full page in one pass (text, HTML tables, LaTeX equations, reading order). Versus Surya 2 it is about 1.7× faster and scores 85.9 on OmniParseBench (best open model under 1B per Datalab).

Blog: Surya 2.5 · Weights: datalab-to/surya-ocr-2.5 · Package: pip install surya-ocr

Key points

1. Speed: ~4.5 pages/s on one H100 at 192 DPI / 32 concurrency (vs ~2.7 for Surya 2+MTP).
2. Accuracy: OmniParseBench 85.9 (gains on text/tables/layout); olmOCR-bench ~83.7; loop-to-cap ~0.1% (vs ~3.8% for Surya 2+MTP).
3. Multilingual: 91-language eval; vocab ~88.5k (emoji + styled math letters); ~2.6× fewer output tokens per page.
4. License: code Apache 2.0; weights modified OpenRAIL-M (research, personal, startups under $5M).
5. Limits: tables as HTML only (no cell boxes); old scans / some languages remain hard.

Why it matters

Competitor sitemap grew a new /blog/surya-2-5 URL with a same-day launch post and HF weights; the Surya line was not previously in radar_leads.

How to try it

pip install surya-ocr   # 0.23.0+ defaults to Surya 2.5
surya_ocr page.pdf --output_dir out
surya_layout page.pdf --output_dir out

1. On NVIDIA the manager starts vLLM (Docker); on Apple Silicon use llama.cpp / Metal.
2. For higher accuracy without self-hosting, use Datalab’s API with Chandra 2.5.
3. GGUF weights: datalab-to/surya-ocr-2.5-gguf.

Keep Surya 2’s table-structure task if you need cell bounding boxes.