v0.20.0

Textualize/textualv0.20.0May 27, 2026by VikParuchuri

AI Summary

Surya OCR 2 introduces a major overhaul with a single 650M-param model handling OCR, layout, and table recognition, replacing the previous multi-model architecture. It introduces new runtime requirements using vllm or llama.cpp and modifies several API endpoints and output schemas.

Key Highlights

State of the art accuracy for its size, achieving 83.3% on olmOCR-bench
Multilingual support covering 91 languages with an 87.2% average benchmark score
High throughput of 5 pages per second on RTX 5090
Unified model approach serving OCR, layout, and table recognition from a single 650M-parameter model

Breaking Changes

SuryaInferenceManager replaces FoundationPredictor
Output schemas changed: text_lines renamed to blocks (with html), layout dropped top_k and added count, table-rec cells dropped is_header/colspan/rowspan
New runtime requirements: Docker + NVIDIA Container Toolkit for GPU or llama.cpp for CPU/Apple Silicon

New Features

Unified 650M-param model for OCR, layout, and table recognition
Integration with vllm and llama.cpp inference backends
Updated output schemas featuring HTML blocks
Separate lightweight torch models for text detection and OCR-error detection

Full Release Notes

# Surya 2 (v0.20.0)

Surya 2 is a ground-up rework: a single **650M-param model** now handles OCR, layout, and table recognition, served by `vllm` (NVIDIA GPU) or `llama.cpp` (CPU / Apple Silicon). Text detection and OCR-error detection remain separate lightweight torch models.

> ⚠️ **This is a major release with breaking API and output-schema changes.** See *Upgrading from v1* below.

## Highlights

- **State of the art for its size** - 83.3% on [olmOCR-bench](https://huggingface.co/datasets/allenai/olmOCR-bench), best in class under 3B params.
- **Multilingual** - 87.2% average across a 91-language internal benchmark.
- **Fast** - 5 pages/s throughput on RTX 5090.

## Breaking changes — upgrading from v1

```python
# v2
from surya.inference import SuryaInferenceManager
from surya.recognition import RecognitionPredictor

manager = SuryaInferenceManager()        # auto-spawns vllm or llama-server
rec = RecognitionPredictor(manager)
predictions = rec([image])
```

- `SuryaInferenceManager` replaces `FoundationPredictor`, and is shared across `LayoutPredictor`, `RecognitionPredictor`, and `TableRecPredictor`.
- **Output schemas changed:** `text_lines` → `blocks` (each with `html`); layout dropped `top_k` and added `count`; table-rec cells dropped `is_header` / `colspan` / `rowspan`.
- **New runtime requirement:** layout / OCR / table-rec need an inference backend - Docker + the NVIDIA Container Toolkit (GPU), or `brew install llama.cpp` (CPU / Apple Silicon). Detection still runs on torch alone.

## Installation

```shell
pip install surya-ocr
```

Then make a backend available (see *Breaking changes* above). Full usage, output schemas, and tuning notes are in the [README](https://github.com/datalab-to/surya#readme).

## Notes

- Try it without installing anything on the [Datalab playground](https://www.datalab.to/playground?utm_source=gh-surya).

**Full Changelog**: https://github.com/datalab-to/surya/compare/v0.17.1...v0.20.0