Surya OCR 2

Datalab's 650M OCR model: text, layout, reading order and tables in 91 languages. Free for research and organizations under US$5M.

In short. Surya OCR 2 reads scans and PDFs, finds each block's layout and reading order, and recovers tables and math, all with one small model of about 650 million parameters. It is widely downloaded, but its weights are free only for research, personal use and organizations under US$5 million.

Summary

Datalab released Surya OCR 2 on 27 May 2026 as a ground-up rework of its Surya toolkit. A single vision-language model, which uses the Qwen3.5 architecture in its configuration, now does full-page OCR, layout analysis with reading order and table recognition, where earlier Surya versions used separate models. Datalab calls it a 650M model, and the checkpoint holds about 686 million parameters in a 1.4 GB BF16 file. It returns each block as HTML with math in KaTeX-compatible tags, labels blocks with 18 layout types such as tables, headers, equations and pictures, and returns rows, columns and cells for tables, with full HTML for spanning cells in an optional mode. Text-line detection stays a separate small PyTorch model.

Datalab reports 83.3% on olmOCR-bench and calls that the best result under 3B parameters. Its table puts Surya OCR 2 just below dots.mocr (3.0B, 83.9) and above LightOnOCR 2-1B (83.2, measured with a different method), and well below Datalab's own Chandra OCR 2 (85.9). The per-source scores show where it is weak: 99.7 on the base tests but only 41.8 on old scans. On Datalab's internal 91-language test it scores 87.2% overall, with 38 languages at 90% or more: English 92.3%, Korean 86.7%, Japanese 86.2%, Chinese 82.5%, Vietnamese 73.2% and Arabic 72.7%. All of these figures are Datalab's own. Throughput is 5.35 pages per second on one RTX 5090 with vLLM at 128 concurrent requests, and about 0.11 pages per second on Apple Silicon with llama.cpp.

Layout, OCR and tables need an inference backend: vLLM on an NVIDIA GPU, which Datalab's release notes pair with Docker and the NVIDIA Container Toolkit, or llama.cpp on a CPU or Apple Silicon. Datalab says the model is built for documents, and photos or natural scenes are not the goal. The code is Apache 2.0, but the weights use a modified Open RAIL-M license: free for research, personal use and organizations with under US$5 million of revenue and under US$5 million raised, with a ban on competing with Datalab and share-alike for derivatives and outputs. In the 30 days to 11 November 2026, Hugging Face counted about 941,000 downloads, a figure that may include automated pulls.

Good for

  • Fast, low-cost OCR with layout, reading order and tables from one 1.4 GB model
  • Machines without a large GPU, since llama.cpp runs it on a CPU or Apple Silicon
  • Many languages, with a published 91-language breakdown to check yours against
  • Research, personal projects and organizations under US$5 million of revenue and funding

Watch out for

  • The weights are free only for research, personal use and organizations under US$5 million of revenue and funding, with a ban on competing products and share-alike for outputs
  • All benchmark figures are Datalab's own, and the 83.3% on olmOCR-bench includes just 41.8 on old scans
  • A vLLM or llama.cpp backend is required, and the v2 API and output format break compatibility with Surya 1
  • It is built for documents, not photos or natural scenes, and Chandra OCR 2 is Datalab's more accurate option

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Surya OCR 2 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Surya OCR 2 BF16 690M 1.59 GB 0.793 GB 0.476 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

The same modified Open RAIL-M license as Datalab's Chandra models, with higher limits. Free for research and personal use, and for commercial use only by organizations with under US$5 million in prior-year revenue and under US$5 million raised in total. It also bans using the model to compete with Datalab and requires derivatives and outputs to carry the same license.

  • Attribution or model name must be shown
  • Free commercial use only below a revenue cap
  • Acceptable-use policy applies

More about OpenRAIL-M (Datalab Surya)

See also

Frequently asked questions

What is Surya OCR 2?
Surya OCR 2 reads scans and PDFs, finds each block's layout and reading order, and recovers tables and math, all with one small model of about 650 million parameters. It is widely downloaded, but its weights are free only for research, personal use and organizations under US$5 million.
Can I use Surya OCR 2 commercially?
Yes, with conditions. Surya OCR 2 is released under the OpenRAIL-M (Datalab Surya), which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Acceptable-use policy applies. Read the license text before you ship a product.
Is Surya OCR 2 open source?
Surya OCR 2 is open-weight, not open source in the strict sense. You can download the model files, but the OpenRAIL-M (Datalab Surya) adds conditions that the Open Source Initiative definition does not allow.
Can I run Surya OCR 2 on my own computer?
Yes. The smallest variant, Surya OCR 2, needs about 0.476 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Surya OCR 2 released?
Surya OCR 2 was released by Datalab on May 27, 2026.
How long a context window does Surya OCR 2 have?
Surya OCR 2 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Chandra OCR 2 Datalab 5.3B 256K OpenRAIL-M (Datalab) (Conditional) 3.66 GB Laptop or phone
LightOnOCR-3 LightOn 853M – 4.54B 256K Apache 2.0 (Permissive) 0.589 GB Laptop or phone
GLM-OCR Z.ai (Zhipu AI) 900M – MIT (Permissive) 0.621 GB Laptop or phone
PaddleOCR-VL Baidu 960M – Apache 2.0 (Permissive) 0.662 GB Laptop or phone