Open-source document intelligence platform Datalab has released its new multilingual OCR model, Surya OCR 2. Despite having only 650 million parameters, it scores 83.3% on the authoritative olmOCR-bench benchmark, ranking first among models under 3 billion parameters. It even outperforms the original 9-billion-parameter version — about 14 times its size — achieving a Pareto optimum in parameter count and accuracy.
Functionally, Surya OCR 2 consolidates layout analysis, text recognition, and table extraction into a single vision-language model (VLM). Independent lightweight models handle text line detection and OCR error detection. A single model call performs full-page OCR and outputs structured HTML with bounding boxes and reading order. Math formulas are output as HTML math tags, and cross-row/cross-column tables are formatted as standard HTML. For multilingual support, the model achieves a 87.2% overall pass rate across 91 languages (82.5% for Chinese), with deep optimizations for damaged documents and handwriting.
On the deployment front, Surya OCR 2 supports two inference backends. On NVIDIA GPUs, it runs in Docker with the vLLM backend, achieving a throughput of 5.35 pages per second on a single RTX 5090. On Apple devices or standard CPUs, it uses llama.cpp to load GGUF format, enabling fully local inference on M1 Macs. The model's source code is open-sourced under Apache 2.0, and the weights are available under OpenRAIL-M for free for individuals, universities, and startups with annual revenue under $5 million. Datalab also offers a paid API with the more powerful 4-billion-parameter Chandra 2 model, including a $5 trial credit.