Skip to main content
Back

Datalab's Surya OCR 2 Beats 9B Models at a Fraction of the Size

Datalab's 650M parameter open-source OCR model matches multi-billion parameter systems while running on a laptop and covering 91 languages.

  • Datalab released Surya OCR 2, a 650M parameter open-source OCR model.
  • Scores 83.3% on olmOCR-bench, best in class under 3B parameters.
  • Handles layout, OCR, and table recognition in a single Qwen3.5-style VLM.
  • Multilingual pass rate of 87.2% across 91 languages, with English at 92.3%.
  • Runs 5.35 pages/s on an RTX 5090 via vllm, 0.108 pages/s on Apple Silicon via llama.cpp.
  • Free for research and startups under $5M revenue via the GitHub repo.

Datalab has released Surya OCR 2, a 650M parameter document OCR model that punches well above its weight class. It reaches 83.3% accuracy on the olmOCR-bench, setting a new bar for models under 3 billion parameters, and it runs on everything from an RTX 5090 to an M1 MacBook.

The pitch is straightforward: one small vision-language model that handles layout analysis, full-page OCR, and table recognition, wrapped in a Python package you can install with pip install surya-ocr. Weights and code sit on Hugging Face and GitHub, with a hosted playground if you'd rather not spin anything up locally.

One VLM to parse them all

The architectural bet here is consolidation. Layout, OCR, and table recognition all share a single vision-language model built on a Qwen3.5-style architecture with about 650M parameters, trained on diverse document images to emit either a layout JSON or a full-page HTML output depending on the prompt. Text-line detection runs as a separate small torch model, a modified EfficientViT segformer trained from scratch on document line annotations.

Practically, that means one model swap between tasks, one set of weights to cache, and one inference server to run. Layout, OCR, and table-rec need an inference backend, either Docker plus the NVIDIA Container Toolkit for GPU or llama.cpp for CPU and Apple Silicon. Detection still runs on torch alone.