Surya 2.5. surya-ocr 0.23.0 uses Surya 2.5 (datalab-to/surya-ocr-2.5) by default: 1.7x as many pages per second as Surya 2 on one H100, and 85.9 on OmniParseBench, the best of any open model under 1B parameters. It almost never loops (0.1% of requests reach the output limit at 192 DPI, against 3.8% for Surya 2 with MTP). See the announcement and the model card.
surya_tablenow defaults to table HTML (--mode full). Surya 2.5 has no row/column geometry; use Surya 2 (SURYA_MODEL_CHECKPOINT=datalab-to/surya-ocr-2) for--mode simple.- New experimental MLX backend for Apple Silicon.
- To stay on Surya 2:
SURYA_MODEL_CHECKPOINT=datalab-to/surya-ocr-2(and, for llama.cpp,SURYA_GGUF_REPO=datalab-to/surya-ocr-2-gguf).
What's Changed
- readme updates by @VikParuchuri in #562
- readme updates by @VikParuchuri in #563
- Surya 2.5 support, release 0.23.0 by @hunter-heidenreich in #568
- Surya 2.5, release 0.23.0 by @hunter-heidenreich in #569
New Contributors
- @hunter-heidenreich made their first contribution in #568
Full Changelog: v0.22.1...v0.23.0