Surya-Ocr-2-Onnx

<p align="center"> <img src="datalab-logo.png" alt="Datalab Logo" width="150"/> </p> <h1 align="center">Datalab</h1> <p align="center"> <strong>State of the Art models for Document Intelligence</strong> </p> <p align="center"> <a href="https://www.datalab.to"><img src="https://img.shields.io/badge/Homepage-datalab.to-blue" alt="Homepage"></a> <a href="https://documentation.datalab.to"><img src="https://img.shields.io/badge/Docs-Read%20the%20docs-blue" alt="Docs"></a> <a href="https://www.datalab.to/playground"><img src="https://img.shields.io/badge/Datalab Playground-Try%20it-orange" alt="Datalab Playground"></a> </p> <hr/>

Surya

Surya is a 650M param OCR model with these features:

• Accuracy - scores 83.3% on olmOCR-bench (top under 3B params)
• Speed - throughput of 5 pages/s on an RTX 5090
• Multilingual - scores 87.2% on an internal benchmark set of 91 languages (more here)
• Layout analysis (table, image, header, etc.) with reading order
• Table recognition (rows + columns)

It works on a range of documents (see usage and benchmarks).

Try Datalab's Managed Platform

Our managed platform runs both Surya, and variants of our highest accuracy model, Chandra.

Get started with $5 in free credits — sign up (takes under 30 seconds) or try our free public playground.

Model Information

<img src="assets/olmocr_size_chart.png" width="700"/>

Surya is named for the Hindu sun god, who has universal vision.

Examples

Commercial usage

The Surya code is licensed under Apache 2.0. The model weights use a modified AI Pubs Open Rail-M license (free for research, personal use, and startups under $5M funding/revenue). For broader commercial licensing of the model weights, visit our pricing page here.

Installation

Install with:

pip install surya-ocr

Usage

Surya 2 runs layout, OCR, and table recognition through a single VLM served by vllm (GPU) or llama.cpp (CPU / Apple Silicon). The inference manager will spawn one for you on first use; you can also point it at an existing server via SURYA_INFERENCE_URL=http://host:port/v1.

• Inspect the settings in surya/settings.py. You can override any setting via env var (e.g. SURYA_INFERENCE_BACKEND=vllm).
• Text detection and OCR errors are separate models.

Interactive App

I've included a streamlit app that lets you interactively try Surya on images or PDF files. Run it with:

pip install streamlit pdftext
surya_gui

OCR (text recognition)

This command will write out a json file with the detected text and bboxes:

surya_ocr DATA_PATH

• DATA_PATH can be an image, pdf, or folder of images/pdfs
• --images will save images of the pages and detected blocks (optional)
• --output_dir specifies the directory to save results to instead of the default
• --page_range specifies the page range to process in the PDF, specified as a single number, a comma separated list, a range, or comma separated ranges - example: 0,5-10,20.

The results.json file contains a dict keyed by input filename (no extension). Each value is a list of page dicts. Each page dict contains:

• blocks - per-block OCR results in reading order
  • label - canonicalized layout label (e.g. Text, SectionHeader, Table, Equation, Picture, Form, PageHeader, ...). See surya/layout/label.py:LAYOUT_PRED_RELABEL for the full canonical-name set.
  • raw_label - original label emitted by the model, before canonicalization
  • reading_order - 0-indexed position in layout output
  • html - block content as HTML (math wrapped in <math>...</math>, tables as <table>...</table>, etc.). "" if the block was skipped
  • polygon - 4-corner polygon in [[x0,y0],[x1,y0],[x1,y1],[x0,y1]] order
  • bbox - axis-aligned [x0, y0, x1, y1] derived from the polygon
  • confidence - mean per-token probability across the block's decode (0-1)
  • skipped - true if the block was a visual label (e.g. Picture) and not OCR'd
  • error - true if the block OCR call failed
• image_bbox - [0, 0, width, height] for the page image

Performance tips

Throughput is governed by the inference backend, not a RECOGNITION_BATCH_SIZE env var. With vllm, raise --max-num-seqs / --max-num-batched-tokens (or SURYA_INFERENCE_PARALLEL on the client side) to keep more pages in flight. With llama.cpp, set SURYA_INFERENCE_PARALLEL to match --parallel on llama-server.

From python

from PIL import Image
from surya.inference import SuryaInferenceManager
from surya.recognition import RecognitionPredictor

manager = SuryaInferenceManager()
recognition_predictor = RecognitionPredictor(manager)

# Default: full-page OCR. One VLM call per page; returns layout + content as
# HTML <div data-bbox=... data-label=...> blocks.
predictions = recognition_predictor([Image.open(IMAGE_PATH)])

# Block mode: pre-run layout, then per-block OCR. Auto-selected when
# `layout_results` is passed.
from surya.layout import LayoutPredictor
layout = LayoutPredictor(manager)
layouts = layout([Image.open(IMAGE_PATH)])
predictions = recognition_predictor([Image.open(IMAGE_PATH)], layouts)

Text line detection

This command will write out a json file with the detected bboxes.

surya_detect DATA_PATH

• DATA_PATH can be an image, pdf, or folder of images/pdfs
• --images will save images of the pages and detected text lines (optional)
• --output_dir specifies the directory to save results to instead of the default
• --page_range specifies the page range to process in the PDF, specified as a single number, a comma separated list, a range, or comma separated ranges - example: 0,5-10,20.

The results.json file will contain a json dictionary where the keys are the input filenames without extensions. Each value will be a list of dictionaries, one per page of the input document. Each page dictionary contains:

• bboxes - detected bounding boxes for text
  • bbox - the axis-aligned rectangle for the text line in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner.
  • polygon - the polygon for the text line in (x1, y1), (x2, y2), (x3, y3), (x4, y4) format. The points are in clockwise order from the top left.
  • confidence - the confidence of the model in the detected text (0-1)
• vertical_lines - vertical lines detected in the document
  • bbox - the axis-aligned line coordinates.
• page - the page number in the file
• image_bbox - the bbox for the image in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner. All line bboxes will be contained within this bbox.

Performance tips

Detection is a torch model. DETECTOR_BATCH_SIZE defaults to an auto-picked value at runtime; override the env var to control VRAM usage on GPU and raise it on larger cards.

From python

from PIL import Image
from surya.detection import DetectionPredictor

det_predictor = DetectionPredictor()
predictions = det_predictor([Image.open(IMAGE_PATH)])

Layout and reading order

This command will write out a json file with the detected layout and reading order.

surya_layout DATA_PATH

• DATA_PATH can be an image, pdf, or folder of images/pdfs
• --images will save images of the pages and detected text lines (optional)
• --output_dir specifies the directory to save results to instead of the default
• --page_range specifies the page range to process in the PDF, specified as a single number, a comma separated list, a range, or comma separated ranges - example: 0,5-10,20.

The results.json file contains a dict keyed by input filename (no extension). Each value is a list of page dicts. Each page dict contains:

• bboxes - layout boxes in reading order
  • polygon - 4-corner polygon [[x0,y0],[x1,y0],[x1,y1],[x0,y1]]
  • bbox - axis-aligned [x0, y0, x1, y1] derived from the polygon
  • label - canonicalized label. One of Caption, Footnote, Equation, ListGroup, PageHeader, PageFooter, Picture, SectionHeader, Table, Text, Figure, Code, Form, TableOfContents, ChemicalBlock, Diagram, Bibliography, BlankPage
  • raw_label - original label emitted by the model
  • position - 0-indexed reading order
  • count - model's token estimate for OCR'ing this block (rounded to multiples of 50; used to size the per-block decode budget)
  • confidence - mean per-token probability across the layout decode (0-1)
• image_bbox - [0, 0, width, height]
• raw - raw JSON the layout model emitted, for debugging
• error - true if the layout call failed

Performance tips

Layout runs through the shared inference backend. Throughput tuning is the same as OCR — see Performance tips above.

From python

from PIL import Image
from surya.inference import SuryaInferenceManager
from surya.layout import LayoutPredictor

layout_predictor = LayoutPredictor(SuryaInferenceManager())
layout_predictions = layout_predictor([Image.open(IMAGE_PATH)])

Table Recognition

This command will write out a json file with the detected table cells and row/column ids, along with row/column bounding boxes. If you want to get cell positions and text, along with nice formatting, check out the marker repo. You can use the TableConverter to detect and extract tables in images and PDFs. It supports output in json (with bboxes), markdown, and html.

surya_table DATA_PATH

• DATA_PATH can be an image, pdf, or folder of images/pdfs
• --images will save annotated row + column overlays alongside the json (optional)
• --output_dir specifies the directory to save results to instead of the default
• --page_range specifies the page range to process in the PDF, specified as a single number, a comma separated list, a range, or comma separated ranges - example: 0,5-10,20.
• --skip_table_detection tells table recognition not to detect tables first. Use this if your image is already cropped to a table.

The results.json file contains a dict keyed by input filename (no extension). Each value is a list of per-table dicts. Each table dict contains:

• rows - detected table rows in reading order
  • polygon / bbox - row geometry (same convention as everywhere else)
  • row_id - 0-indexed row id
• cols - detected table columns
  • polygon / bbox - column geometry
  • col_id - 0-indexed column id
• cells - geometric row × column intersections (simple mode)
  • polygon / bbox - cell geometry
  • row_id, col_id, cell_id
• html - full <table>...</table> HTML (only populated when predict_full is used; handles spanning cells / header rows). null in simple mode.
• mode - "simple" or "full"
• image_bbox - the table crop bbox
• error - true if the table_rec call failed
• raw - raw model output, for debugging

Performance tips

Table recognition routes through the shared VLM. Throughput tuning is the same as OCR.

From python

from PIL import Image
from surya.inference import SuryaInferenceManager
from surya.table_rec import TableRecPredictor

table_rec_predictor = TableRecPredictor(SuryaInferenceManager())

# Default: rows + columns only, cells derived from intersections.
table_predictions = table_rec_predictor([Image.open(IMAGE_PATH)])

# Or full HTML output (better for spanning cells / headers):
# table_predictions = table_rec_predictor.predict_full([image])

Math / equations

Surya 2 handles math inline as part of full-page OCR — recognized equations come back inside <math>...</math> tags in the same HTML output as surrounding prose, in KaTeX-compatible LaTeX. No separate LaTeX OCR pass.

Inference Backends

Layout / OCR / table_rec all share one VLM, served either by vllm (GPU) or llama.cpp (CPU / Apple Silicon). The SuryaInferenceManager will spawn one automatically; you can also point at a pre-running server:

# Attach to an existing vllm
export SURYA_INFERENCE_BACKEND=vllm
export SURYA_INFERENCE_URL=http://localhost:8000/v1

Limitations

• This is specialized for document OCR. Performance on photos or natural scenes is not the goal.
• Layout / OCR / table_rec all need a running inference backend (vllm or llama.cpp). Detection runs purely on torch and works without it.

Troubleshooting

If OCR isn't working properly:

• Try increasing resolution of the image so the text is bigger. If the resolution is already very high, try decreasing it to no more than a 2048px width.
• Preprocessing the image (binarizing, deskewing, etc) can help with very old/blurry images.
• You can adjust DETECTOR_BLANK_THRESHOLD and DETECTOR_TEXT_THRESHOLD if you don't get good results. DETECTOR_BLANK_THRESHOLD controls the space between lines - any prediction below this number will be considered blank space. DETECTOR_TEXT_THRESHOLD controls how text is joined - any number above this is considered text. DETECTOR_TEXT_THRESHOLD should always be higher than DETECTOR_BLANK_THRESHOLD, and both should be in the 0-1 range. Looking at the heatmap from the debug output of the detector can tell you how to adjust these (if you see faint things that look like boxes, lower the thresholds, and if you see bboxes being joined together, raise the thresholds).

Manual install

If you want to develop surya, you can install it manually with uv:

git clone https://github.com/datalab-to/surya.git
cd surya
uv sync --group dev      # installs runtime + dev deps
uv run surya_ocr ...     # or `uv shell` to enter the venv

Benchmarks

Surya 2 is a single VLM that handles layout analysis, OCR (full-page or per-block), and table recognition in one model. We evaluate end-to-end on olmOCR-bench — the standard quality benchmark for document parsers.

olmOCR-bench

Pareto-optimal, and best in class under 3B params.

* LightOnOCR 2-1B uses a different benchmark methodology than the other entries (see their release notes); the score is included for context but is not directly comparable.

Comparison scores from the olmOCR-bench dataset card.

Surya 2, per-source pass rate on the default preset (8,413 tests total):

Multilingual

We also evaluate Surya 2 against a 91-language internal benchmark covering text accuracy, layout, tables, math, and reading order in documents drawn from each language.

Overall pass rate: 87.2% across 91 languages. 38 of the 91 languages score ≥ 90%; 76 score ≥ 80%.

Top 15 widely-spoken languages:

See https://github.com/datalab-to/surya/blob/master/static/docs/multilingual.md for the full 91-language table.

Throughput

Full-page OCR, 96 DPI input (~2,400 output tokens/page average), measured client-side against a running inference server.

RTX 5090 (vllm)

vllm/vllm-openai:v0.20.1, single RTX 5090 (32 GB).

Apple Silicon (llama.cpp / Metal)

llama-server with Metal backend.

Reproducing

We score Surya 2 on olmOCR-bench by serving the model with vllm (or llama.cpp) and running the olmOCR-bench harness from allenai/olmocr, with some adjustments applied to account for our output HTML format.

Training

Layout, OCR, and table recognition all share a single vision-language model (Qwen3.5-style architecture, ~650M params). It's trained on diverse document images to emit either a layout JSON or a full-page HTML output, depending on prompt. Text-line detection is a separate small torch model — a modified EfficientViT segformer trained from scratch on document line annotations.

If you want help finetuning Surya on your own data, or to use our managed training stack, reach us at [email protected].

--

surya-ocr-2 → ONNX (ONNX Runtime GenAI)

Convert datalab-to/surya-ocr-2 — an OCR VLM on a Qwen3.5-VL backbone (model_type: qwen3_5) — to three ONNX sub-parts (text decoder, vision encoder, text embedding) with Olive, re-composed at runtime by ONNX Runtime GenAI.

Same hybrid GatedDeltaNet architecture as Qwen-Qwen3.5-0.8B/builtin/; not the old chandra/builtin recipe (different architecture, Qwen3-VL-8B). optimize.py is fully Python-driven — no static Olive JSON configs. See HANDOFF.md for the full history: the tokenizer/bos_token_id fixes, the real task prompts, and the traced cuda_fp32 limitation.

Build

python optimize.py --device cpu   --precision int4    # all three, cpu — BUILT + TESTED
python optimize.py --device cpu   --precision fp32    # BUILT + TESTED
python optimize.py --device cuda  --precision fp16    # gpu, fp16 (recommended for teamspace)
python optimize.py --device cuda  --precision int4
python optimize.py --device webgpu --precision fp16

Valid (device, precision): cpu×{int4,fp32}, cuda×{int4,fp16}, webgpu×{int4,fp16} (fp16 is GPU-only — Olive's ModelBuilder pass rejects it on CPU. cuda_fp32 is deliberately excluded — architecturally broken for this model, not just untested: genai's builder resolves (cuda, fp32) to the packed-Attention op, which this model's attention override doesn't support — KeyError('root_input'). Traced through genai's own builder source; see HANDOFF.md §5.1). Builds from a local ./model snapshot if present, else downloads datalab-to/surya-ocr-2. Output → <device>_<precision>/models/ (text.onnx, vision.onnx, embedding.onnx, genai_config.json, processor_config.json, tokenizer files).

Windows: run with PYTHONUTF8=1 (a vision-export print crashes on cp1252 otherwise).

Inference and eval

inference.py drives one image through a built ONNX target using surya's own real task prompts (sourced from datalab-to/surya's surya/inference/prompts.py — see tasks.py; don't invent prompts for this model, a paraphrase can look right and still be wrong):

python inference.py --model_path cpu_int4/models --task ocr --image model/assets/corporate.png
python inference.py --model_path cpu_int4/models --task layout --image model/assets/form.png

eval.py batch-evaluates one or more built targets against the PyTorch checkpoint on real document pages (validates layout/table_rec output against surya's actual JSON schema, and reports text similarity vs. PyTorch for every task):

python eval.py --model model cpu_int4/models cpu_fp32/models   # explicit targets
python eval.py --model model                                    # auto-discovers built targets

Use the unsuffixed pages in model/assets/ (corporate.png, form.png, handwritten.png, ...) as input — the *_layout/_text/_reading/_tablerec.png siblings are this README's own reference-output visualizations, not inputs.

Verified end-to-end (2026-07-30)

Both cpu_int4 and cpu_fp32 built successfully and were tested with real inference across all three tasks (layout, table_rec, ocr) against real document pages — 12/12 runs clean, valid schema-conforming output, cross-checked against the PyTorch checkpoint. This is also the direct regression test for the KeyError('root_input') failure this recipe was written to avoid.

Three runtime bugs were found and fixed along the way (see HANDOFF.md §4 and §7.4):

• bos_token_id: null crashing genai's config parser (this model has no real BOS token).
• A WordLevel tokenizer genai's C++ runtime can't load (re-encoded as BPE with an empty merge list — provably equivalent here since the pre-tokenizer already isolates single characters).
• A U+FFFD (replacement character) vocab entry that genai fails to index into its reverse lookup map, breaking decode of any generated token with that id — surfaces only on long free-form output (ocr, full-page HTML). Fixed with a Python-side fallback decoder (inference.py::load_id_to_token) used only when genai's own decode call fails.

Note: --max_length needs headroom for the image too — a single page can produce 2500–7000+ input tokens depending on resolution.

--

Thanks

This work would not have been possible without amazing open source AI work:

• Qwen3-VL from Alibaba
• vllm and llama.cpp for inference
• Segformer from NVIDIA
• EfficientViT from MIT
• timm from Ross Wightman
• transformers from huggingface
• Olive from Olive
• CRAFT, a great scene text detection model

Thank you to everyone who makes open source AI possible.

Citation

If you use surya (or the associated models) in your work or research, please consider citing us using the following BibTeX entry:

@misc{paruchuri2025surya,
  author       = {Vikas Paruchuri and Datalab Team},
  title        = {Surya: A lightweight document OCR and analysis toolkit},
  year         = {2025},
  howpublished = {\url{https://github.com/datalab-to/surya}},
  note         = {GitHub repository},
}