Datalab's Surya OCR 2.5 Runs 1.7x Faster by Cutting Output Tokens 61%
Datalab's new 689M-parameter OCR model hits 85.9 on OmniParseBench, runs 1.7x faster than its predecessor, and ships with character-level tokenization.
Takeaways
- Datalab released Surya OCR 2.5, a 689M-param VLM scoring 85.9 on OmniParseBench.
- Runs 1.7x faster than Surya 2, hitting 4.5 pages/s on one H100 at 192 DPI.
- Writes about one-third the output tokens per page compared to Surya 2.
- Loops 38x less often than Surya 2 on hard sets, with character-level tokenization.
- Covers 91 languages; 52 score above 80% but some regressions in English and Russian.
- Drops row and column table geometry; use Surya 2 checkpoint if needed.
Datalab has released Surya OCR 2.5, a 689 million-parameter vision-language model for document layout analysis, full-page OCR, and table recognition. Datalab reports the highest parsing accuracy among open OCR models below one billion parameters, plus 1.7 times the throughput of Surya 2 on its published H100 test.
The release targets developers who need structured document output without serving a multi-billion-parameter model. It produces JSON for layout analysis and HTML for OCR and tables, including merged cells and KaTeX-compatible LaTeX for mathematical content.
One VLM, three outputs
Surya 2.5 uses a Qwen3.5-style vision-language architecture trained on diverse document images. A prompt selects the task and output format:
- Layout analysis: structured JSON describing page regions.
- Full-page OCR: HTML containing text, structure, and mathematical notation.
- Table recognition: HTML tables with
rowspanandcolspanfor merged cells.
All three tasks share the same VLM backend, served through vLLM on NVIDIA GPUs or llama.cpp on CPUs and Apple Silicon. Text-line detection uses a separate, smaller PyTorch model: a modified EfficientViT SegFormer trained from scratch to locate lines in document images.
Accuracy rises as output shrinks
Datalab’s published results show gains on two document-parsing benchmarks that evaluate recovered content and structure:
| Benchmark | Surya 2.5 | Surya 2 | Change |
|---|---|---|---|
| OmniParseBench | 85.9 | 83.3 | +2.6 points |
| olmOCR-bench | 83.7 | 82.8 | +0.9 points |
The model card reports improvements across text, tables, and layout. On one H100 at 192 DPI with vLLM,