Datalab releases Surya OCR 2.5, 689M-parameter document VLM

Datalab released Surya OCR 2.5, what the model card calls a "689M parameter Vision Language Model" for document OCR, layout analysis, and table recognition. It scores 85.9 on OmniParseBench, up from 83.3 for Surya 2, and 83.7% on olmOCR-bench against 82.8% for the prior version. On an H100 running vLLM at 96 DPI the model clocks 6.19 pages per second versus 2.28 for Surya 2; at 192 DPI, the card calls it "1.7x faster than Surya 2 with MTP."

The single VLM covers OCR, layout analysis, table recognition, and reading order in one pass, with 91 languages supported. Twenty-one score at or above a 90% pass rate, 52 at or above 80%. The biggest benchmark jump came on tables, from 74.3 to 78.3 on OmniParseBench.

Code ships under Apache 2.0, with model weights released under OpenRAIL-M, which the card describes as free for research, personal use, and startups under $5M funding/revenue; larger users route through paid licensing from Datalab. Two of the researchers we follow posted the model page the day it went up.