Surya OCR 2: 650M Params, 91 Languages, State-of-the-Art Under 3B

Everyone is sleeping on this new OCR model!
(open-source, 91 languages, < 1B params)
Datalab just released Surya OCR, which is just a state-of-the-art OCR model scoring 83.3% on the olmocr benchmark (top under 3B)
- 650M params
- supports 91 languages
- 5 pages/s on RTX 5090
- runs on CPU, GPU, MPS
- 83.3% olmocr bench score (top under 3B)
- Full layout information
- Extracts + captions images and diagrams
- Strong handwriting, math, form, table support

Every single spec matches Datalab's own official announcement almost exactly:

• 650M params, 83.3% olmOCR-bench, top under 3B — confirmed exactly. Surya OCR 2 is a 650M-parameter open-source OCR model that scores 83.3% on olmOCR-bench, hits 87.2% on a 91-language multilingual eval, and runs on CPU, GPU, and MPS.
• 91-language support — confirmed. Surya 2 gets as much performance as possible out of a 650M param model while staying competitive across the olmOCR benchmark and 91 languages, achieved through architectural modification and training data mix.
• 5 pages/s on RTX 5090 — confirmed almost exactly (official figure is 5.35 pages/s, post rounds to 5). It offers high deployment efficiency, achieving 5.35 pages per second on NVIDIA GPUs and supporting local inference on Apple M1 devices.
• Runs on CPU/GPU/MPS, unified layout+OCR+table model — confirmed. Layout, OCR, and table recognition all share a single vision-language model (Qwen3.5-style architecture, ~650M params), trained to emit either a layout JSON or a full-page HTML output.
• Open-source — confirmed, though with nuance: code is Apache 2.0, weights are modified OpenRAIL-M (free for research, personal use, and startups under $5M). So it's open-source code with a somewhat restricted weights license, not a fully permissive Apache/MIT weights release — worth knowing if you're using it commercially at scale.
• Handwriting, math, forms, table support — confirmed as core improvements. Surya OCR 2 improves accuracy significantly across tables, handwriting, forms, math, and layout compared to its predecessor.