Datalab Ships Surya 2.5, the Fastest Open-Weight Document Parser Under 1 Billion Parameters
The 689M-parameter OCR model posts a vendor-reported 85.9 on OmniParseBench and runs 1.7x faster than Surya 2, underscoring the shift toward small, high-accuracy parsing on modest hardware.
Datalab has released Surya 2.5, an open-weight OCR and document parsing model with 689 million parameters. The company reports 4.5 pages per second on a single H100 — 1.7x faster than Surya 2 — and a score of 85.9 on OmniParseBench, the highest among open models under 1B parameters. Table recognition rose to 78.3, repetitive-loop failures dropped from 3.8% to 0.1% of requests, and each page now needs 2.6x fewer output tokens since the model can now write text. Coverage spans 91 languages, with 21 scoring at or above 90.
The code ships under Apache 2.0 while the weights carry the OpenRAIL-M license, free for research, personal use and startups below $5 million in revenue. Datalab flags real limits: no table bounding-box coordinates, weak results on old or poor-quality scans (44.3 on the olmOCR-bench legacy subset), and Arabic at just 45.4%.
The release lands amid a wave of sub-1B parsers including Baidu Unlimited OCR 2, MiniCPM-OCR and Indra 1.1, all pushing document parsing toward commodity, CPU-capable hardware — a direction that lowers per-page inference costs for enterprises running on-premise pipelines. All benchmarks remain vendor-reported, and Indonesian has not been publicly tested.