AI Briefing
KOSign in

Datalab releases Surya OCR 2.5, a 689M parameter model with 85.9 OmniParseBench score

·2026.10.31 02:55

Key point

The 689M parameter model supports 91 languages and processes documents at 4.5 pages/s on a single H100 GPU.

Details

  • Datalab releases Surya OCR 2.5, a 689M parameter model scoring 85.9 on OmniParseBench and 83.7% on olmOCR-bench.
  • Throughput reaches 4.5 pages/s on one H100 at 192 DPI, representing a 1.7x improvement over Surya 2 with MTP.
  • Multilingual performance scores 74.9% on an internal benchmark covering 91 languages.
  • Code is licensed under Apache 2.0, while model weights use a modified AI Pubs Open Rail-M license, free for research and startups under $5M funding/revenue.
  • Table extraction outputs HTML with merged cells, but row and column geometry requires a Surya 2 checkpoint.
  • Inference for layout, OCR, and table recognition requires a backend such as vllm or llama.cpp.
ocrdocument-aiopen-sourcehugging-facesuryadatalab
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditNVIDIA Releases AlpaJudge Learned Judge Models for Autonomous Driving ReasoningHuggingFace BlogHugging Face Releases AnySpeechRecognizer and AnyDecisionModel Swift Packages for Apple PlatformsRedditOpenBMB releases MiniCPM-OCR 1.0 and MiniCPM-OCR-Jev for local document parsing and verificationPyTorchKRMicrosoft and ThinkingBox-Training Released: Qwen3.8-27B Repeat Success Rate Improved 2.8xHacker NewsJetBrains releases ThinkRail, an open-source agentic IDE built on git worktrees and spec graphsHacker Newsspinf-serve: Open-source inference server for bulk LLM scoring on single GPU