We benchmarked 17 open OCR models on OmniParseBench, our new OCR benchmark that builds on AllenAI's olmOCR-bench.
8
9
155
5,351
It's always a little suspect when someone has their own models benchmarked at the top.
Every adapter is in the repo, and scores can be audited here - github.com/datalab-to/omn… and huggingface.co/datasets/datal… .
Here's a per-category breakdown.
Every adapter is in the repo, and scores can be audited here - github.com/datalab-to/omn… and huggingface.co/datasets/datal… .
Here's a per-category breakdown.
1
0
1
542
For what it's worth, these scores generally align with our vibe tests. dots.mocr and MinerU are quite strong.
0
0
2
333
OCR is now 50 cents for 1k pages. Introducing Datalab lite mode:
- 20-30x cheaper than Azure, Google and AWS layout APIs
- 2x faster on long documents
- Wins on text accuracy
- 20-30x cheaper than Azure, Google and AWS layout APIs
- 2x faster on long documents
- Wins on text accuracy
14
27
326
12,024
On OmniParseBench (documents without math), lite scores 75.1 vs 42.3 for Azure Read, and beats Azure Layout on text.
It's far ahead on non-Latin scripts: 97% on Armenian, 97% Ethiopic, 95% Kannada. Azure Layout: 21%, 14%, 9%.
It's far ahead on non-Latin scripts: 97% on Armenian, 97% Ethiopic, 95% Kannada. Azure Layout: 21%, 14%, 9%.
1
0
2
680
Limitations: this will not work well on handwriting or complex layouts, including tables.
We're continuing to improve lite mode, please share any feedback.
We're continuing to improve lite mode, please share any feedback.
0
0
2
523
How do you know if your OCR is correct?
Verify it with our new word confidence model. See exact positions and confidence values of words, even on messy scans.
Verify it with our new word confidence model. See exact positions and confidence values of words, even on messy scans.
15
17
268
9,332
Read more at datalab.to/blog/word-boxe… , or try it at datalab.to/playground (check word bboxes).
1
3
14
1,049
Indra OCR 1.1 is out now!
- 1M param main OCR model, 44 pages/s on CPU
- Within 4 points of 3.4B param DeepSeek OCR 2 on OmniParseBench
- 15x faster than Tesseract, and more accurate
- 80+ languages, plus handwriting
- 1M param main OCR model, 44 pages/s on CPU
- Within 4 points of 3.4B param DeepSeek OCR 2 on OmniParseBench
- 15x faster than Tesseract, and more accurate
- 80+ languages, plus handwriting
73
351
3,005
225,285
The main changes from 1.0 are improved language support, handwriting (this is limited and mostly English for now), and fixes for different rotation and layout issues.
1
0
24
5,019
I hope it works well for you. We're continuing to improve Indra OCR - let us know if you have any feedback.
6
1
23
4,032
This week, we launched Indra, a 1M parameter OCR model that's 15x faster than Tesseract on CPU and much more accurate.
I still can't believe how good this model is.
I still can't believe how good this model is.
109
242
2,943
104,870
Why do VLMs sometimes output 罪 instead of "vain"?
An investigation into script confusions, and how we fixed them 🧵
An investigation into script confusions, and how we fixed them 🧵
3
21
210
21,409
We used a novel training technique to fix this in Chandra 2.5, live on our API now. On a hard test set, it has 7x fewer script confusions than Chandra OCR 2.4, and fewer than Gemini 3.8 Flash or Claude Opus 5.5.
1
0
15
842