Is Surya OCR 2.5 state of the art?Pending benchmarks.Announced Oct 29, 2026; no independent results yet.
Datalab's open-weights OCR model: one 689M-parameter vision-language model reads a full page and returns each block's type, position and content (text, HTML tables, LaTeX equations) in reading order. It uses character-level tokenization with some merged tokens. Weights on Hugging Face under a modified OpenRAIL-M license (free for research, personal use and startups under $5M); code Apache-2.0.
The details so far
- Lab
- Datalab
- Announced
- Oct 29, 2026
- API
- Not yet
- Open weights
- Published · Hugging Face (datalab-to/surya-ocr-2.5), plus a GGUF build (datalab-to/surya-ocr-2.5-gguf)
- Parameters
- 689M (689,356,864, BF16)
- Type
- Document OCR, layout and reading order in one vision-language model (Qwen3.5-style architecture, per the model card)
- License
- Weights: modified AI Pubs OpenRAIL-M (free for research, personal use and startups under $5M funding/revenue; broader commercial use needs a Datalab license). Code: Apache-2.0
- Tokenization
- Character-level, with about 20,000 merged multi-character tokens added; 88,526-token vocabulary
- Package
- surya-ocr 0.23.0 and later use Surya 2.5 by default (runs on vLLM or llama.cpp)
- Limitations
- No table cell coordinates (tables come back as HTML only); per Datalab, use Surya 2 for table structure boxes
- Previous version
- Surya OCR 2
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 85.9 on OmniParseBench (Datalab's own benchmark, vendor-reported), up from 83.3 for Surya 2, and the highest Datalab has scored among open models under 1B parameters.
Lab claim: 83.7 on olmOCR-bench (Datalab's run), up from 82.8 for Surya 2.
Lab claim: 1.7x faster than Surya 2 with MTP, 4.5 pages per second on one H100 at 192 DPI.
Lab claim: loops 38x less than Surya OCR 2 on a hard set (0.1% of requests hit the output cap, against 3.8%).
Lab claim: 52 of 91 languages score 80% or above on Datalab's internal multilingual benchmark (74.9% overall pass rate).
Sources
LABDatalab blog: Surya 2.5datalab.to/blog/surya-2-5 · Oct 29, 2026LABVik Paruchuri (Datalab founder) on Xx.com/VikParuchuri/status/2115904821136114003 · Oct 29, 2026REPOdatalab-to/surya-ocr-2.5 on Hugging Facehuggingface.co/datalab-to/surya-ocr-2.5REPOdatalab-to/surya on GitHub (v0.23.0)github.com/datalab-to/surya/releases/tag/v0.23.0 · Oct 29, 2026Tracking continues until an independent board lists it.