#1313th of 82 in Data and scraping for AIFor data and documents, documents and ocr
surya
Document OCR, layout and table recognition VLM with a lightweight Python client
- Stars
- 21.6k
- License
- Apache-2.0
- Last commit
- Nov 2026
- Last release
- Nov 2026
- Language
- Python
19,515 on 28 Apr 202621,606 on 17 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Datalab is hiring: 4 open roles
Overview
Surya is a 689M-parameter vision-language model that does page OCR, layout analysis with reading order, and table recognition (HTML with colspan/rowspan). The surya-ocr package is a thin client that spawns or attaches to a vllm, llama.cpp or MLX server. It ships CLI commands, a Python API and a local web viewer (surya_gui). Code is Apache-2.0; model weights use a modified OpenRAIL-M license.
Who it is for: Engineers needing self-hosted OCR, layout and table extraction
Strengths
- One VLM handles OCR, layout, reading order, tables and inline math (KaTeX LaTeX)
- Backends: vllm on NVIDIA, llama.cpp on CPU, mlx_vlm on Apple Silicon
- Reports 85.9 on OmniParseBench and 4.5 pages/s on one H100 at 192 DPI
- Can attach to any OpenAI-compatible server via SURYA_INFERENCE_URL
Weaknesses
- Model weights are not Apache: free only for research, personal use, and startups under $5M
- NVIDIA GPU path needs Docker or Podman plus the NVIDIA Container Toolkit
- Outputs blocks and HTML only; full Markdown or RAG conversion needs the separate Marker project
- Self-hosted servers must return special tokens (llama-server needs --special)
What it needs
- GPU optional
- Needs vllm (Docker/Podman + NVIDIA Container Toolkit), llama.cpp llama-server, mlx_vlm
- Models: Surya 2.5 (689M VLM), OpenAI-compatible servers
- port 8000
What changed
From the release notes. No flag means the note does not say so, not that the upgrade is safe.
Known problems
Open now, reported in the last 90 days
Titles are the reporters' own words, taken from the project's GitHub issues and ranked by reactions. We have not tried to reproduce any of them.
Also in Data and scraping for AI
See all 82| Rank | Project | Score |
|---|---|---|
| 1 | opendataloader-pdfPDF parser that outputs Markdown, JSON with bounding boxes, and Tagged PDF | 92 out of 100 |
| 2 | mineruParses PDFs, images and Office files into Markdown and structured JSON | 91 out of 100 |
| 3 | doclingPython library that converts PDFs, Office files and more into structured text | 91 out of 100 |
| #4 | agent-reachCLI that gives coding agents read access to login-gated and Chinese web platforms | 88 out of 100 |
| #5 | markitdownPython library and CLI that converts office files, PDFs and audio to Markdown | 87 out of 100 |