Datalab's Surya 2 Beats 9B Models at OCR With Just 650M Parameters
Datalab's new 650M parameter open-source OCR model scores 83.3% on olmOCR-bench, beating every system under 3B params while running on a MacBook.
Takeaways
- Datalab released Surya 2, a 650M param open-source OCR model
- Scores 83.3% on olmOCR-bench, best in class under 3B parameters
- Single VLM handles layout, OCR, and table recognition in one model
- Covers 91 languages with 87.2% overall pass rate on internal eval
- Runs 5.35 pages/s on RTX 5090, also works on CPU and Apple Silicon
- Apache 2.0 code, weights free for research and startups under $5M
Document parsing has been stuck in an awkward middle ground. The accurate models are huge multi-billion parameter VLMs that need serious GPUs, while the lightweight options miss languages, mangle tables, and choke on math. Datalab's Surya 2 tries to collapse that tradeoff with a single 650M parameter model that handles OCR, layout analysis, and table recognition in one shot, and runs anywhere from a 5090 to an Apple Silicon laptop.
The repo has crossed 20k stars and the new release puts it in a interesting spot on the size-accuracy curve. Surya OCR 2 hits 83.3% on the olmOCR benchmark, pareto-optimal and competitive with models 5 to 50x larger while being small enough to run on a MacBook.
One VLM doing four jobs
The architectural bet here is consolidation. Layout, OCR, and table recognition all share a single vision-language model with a Qwen3.5-style architecture at around 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 the one exception, kept as a separate small torch model, a modified EfficientViT segformer trained from scratch on document line annotations.
The practical effect is that you spin up one inference backend and route layout, OCR, and table calls through it, instead of juggling three separate models. Math is folded in as well: recognized equations come back inside