Indra OCR 1.1: Is a GPU-Free, 1-Million-Parameter OCR Worth It?
Indra OCR 1.1 runs without a GPU on a model of about 1 million parameters; we explain its speed, accuracy, language support and license terms with numbers.
Contents 12
Indra OCR 1.1 is the new version of Datalab's OCR engine, built around a main model of roughly 1 million parameters, and it runs on CPU alone with no GPU required. It was released on October 19, 2026, and adds handwriting, support for 80+ languages and rotated pages. The code is free; the model weights come with license conditions.
In short:
- Indra OCR 1.1 is an OCR engine of about 1 million parameters that needs no GPU.
- It reaches about 44 pages/second on 32 cores and about 210 ms/page on 4 cores.
- Its handwriting score rose from 8.3 to 25.2, but Datalab itself still calls that weak.
- The model weights are free for startups with less than $20 million in funding or revenue.
What is Indra OCR and what is it for?
OCR is the job of turning the text in an image or scanned document into text a machine can read. Any business that handles invoices, contracts, forms or archive documents runs into this need sooner or later. Indra OCR is a small, fast engine built by Datalab for exactly this job.
Its main model has about 1 million parameters. That is tiny next to large vision-language models, and the payoff is immediate: the model runs on ordinary server CPUs without a GPU. Setup is simple too: the command pip install indra-ocr is enough, and Python 3.10 or later is required. Details are on the PyPI page and in Datalab's release post.
What changed in version 1.1?
The release came on October 19, 2026, after version 0.3 on October 13. According to Datalab's blog post and a news summary, the main changes are:
- Handwriting recognition
- Support for 80+ languages
- Reading of rotated pages
- Higher heading, table and text scores on OmniParseBench
Below we compare the numbers one by one.
How much better is it than 0.3?
OmniParseBench is a test set that measures document parsing quality. The heading score went from 62.3 in 0.3 to 66.7 in 1.1. With the --math option, that score rises from 67.0 to 71.4. The table score also climbed from 56.3 to 61.1.
| Indra OCR 1.1 (--math) | 71,4 |
|---|---|
| Indra OCR 1.1 | 66,7 |
| Indra OCR 0.3 (--math) | 67,0 |
| Indra OCR 0.3 | 62,3 |
Kaynak: Datalab
Overall, each version step brings a gain of 4-5 points. On top of that, two specific capabilities saw bigger jumps:
- Rotated pages: The score rose from 25.3% to 50.3%. In other words, reading quality nearly doubled even when the page is skewed or scanned sideways.
- Handwriting: The score rose from 8.3% to 25.2%.
How reliable is the handwriting support?
Short answer: not yet. The handwriting score tripled, but Datalab itself describes the result as "still weak." A score of 25.2% is not enough for a system meant to process handwritten forms or notebooks automatically.
So for now, handwriting should be treated as a bonus feature. If you plan to put handwriting recognition at the center of your production pipeline, do not decide without testing on your own sample documents. Printed text and tables are far more mature.
How fast is Indra OCR?
Speed is the most striking thing about this engine. The official figures on the PyPI page are:
- About 44 pages/second on 32 CPU cores
- About 210 ms/page on 4 cores
- OmniParseBench median time: 212 ms in auto mode, 344 ms in OCR-only mode, 357 ms with
--math
The Tesseract comparison is on the same page: on a single thread, Tesseract spends 3.4 seconds per page. In the developer's announcement, a 15x speed advantage over Tesseract is the headline claim. Someone who wanted to test it independently shared that they ran the full benchmark on their own Mac Studio and that accuracy matched the announced figure to within about 0.2 points, but the speed advantage over Tesseract stayed around 3x. The gap may vary with hardware and settings. Do not rely on a big speed promise without measuring on your own server.
How does it compare with competitors on accuracy?
In the official comparison, Indra OCR is ahead of the open-source alternatives:
On the heading score, Indra OCR also leads with 66.7 points, ahead of the full PaddleOCR PP-StructureV3 pipeline at 58.4. And a model of 1 million parameters earns these scores. In text recognition it scores more than double what Tesseract does.
The lesson here is that model size alone is not decisive. According to the developer's announcement, the main model trails the OmniParseBench score of the 3.4-billion-parameter DeepSeek OCR 2 by only about 4 points. The size difference is roughly three thousand to one.
Which languages does it read?
The README reports measurements for 95 languages. Of these, 83 have a character F1 score of at least 0.90, and 67 have at least 0.95. Character F1 measures how many of the recognized characters are correct; above 0.95 is considered reliable for most jobs.
The sources we have do not say explicitly whether Turkish is on this list. If you will process Turkish documents, run a small validation test with your own samples, especially on characters like ı, İ, ş and ğ. We covered how these characters can break software in our post on the Turkish İ problem.
License: who can use it for free?
The license has two parts:
- Code: Free.
- Model weights: Free for research, personal use, and startups with less than $20 million in funding or revenue.
This covers most small and medium-sized businesses. But if you are a growing company, you may need a commercial license once you cross the threshold. Write this condition into your product plan from the start. If you are interested in similar support offered to startups, you can also see our startup support programs guide.
Who is Indra OCR suited for?
It is a sensible candidate when:
- Your team does not want to pay for GPUs and processes documents on ordinary servers.
- You process printed documents such as invoices, forms and contracts in bulk.
- You also want to extract table and heading structure; the model provides line positions and reading order as well.
- You want a solution that is more accurate than Tesseract but not as expensive as cloud OCR services.
Be cautious when:
- Handwriting is central to your work; a 25.2% score is not enough.
- With very complex, irregular documents, large models may still give better results.
- You are approaching the license threshold; clarify the commercial terms.
Frequently asked questions
Does Indra OCR require a GPU?
No. The main model, at about 1 million parameters, runs on CPU. It delivers about 210 ms per page on 4 cores and about 44 pages/second on 32 cores.
How do I install Indra OCR?
Install it with pip install indra-ocr. Python 3.10 or later is required. The current version, 1.1.2, is published on PyPI.
Can it read handwriting reliably?
Not yet. Although the handwriting score rose from 8.3% to 25.2%, Datalab describes the result as "still weak." It is much stronger on printed text.
Can I use it in a commercial project?
The code is free. The model weights are free for startups with less than $20 million in funding or revenue. Companies above that threshold should confirm the license terms with Datalab.
Conclusion
Indra OCR 1.1 is a strong counterexample to the assumption that good OCR needs a big model: 1 million parameters, GPU-free operation and scores ahead of open-source rivals. Handwriting is still immature, but for printed documents and tables it is a serious option. Before deciding, the safest route is to run a small trial with your own documents.
At EngerekTech, we support teams with document processing, OCR integration and custom business software projects. If you want help choosing the right tool for your own documents, get in touch with us.
Sources
