StudioDatalab releases Surya 2.5, a 689M-parameter OCR model handling layout, OCR and tables in one VLM
Original title: datalab-to/surya-ocr-2.5 · Hugging Face
Datalab has released Surya 2.5, an open-source document OCR model with 689M parameters that runs layout analysis, full-page OCR and table recognition through a single VLM. It scores 85.9 on OmniParseBench versus 83.3 for Surya 2 and reaches 4.5 pages/s on one H100 at 192 DPI, about 1.7x Surya 2 with MTP. The code is Apache 2.0, weights are free for research and personal use, and inference runs locally via vllm, llama.cpp or an experimental MLX backend; table row/column geometry still requires a Surya 2 checkpoint.
The official model card provides full benchmark data, throughput comparisons and local deployment paths, letting readers assess gains over existing document parsing pipelines.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.