Datalab · Released, not on our provider
Indra OCR (2026): what Datalab’s CPU OCR model is, how fast it reads, what the license allows and how to run it
Indra OCR is Datalab’s open OCR model, first released on October 13, 2026, that turns PDFs and images into markdown, text or JSON with boxes on a plain CPU with about 1M recognition parameters, scores 66.7 on Datalab’s own OmniParseBench, and can be downloaded under a modified OpenRAIL-M license that limits large companies.
Use it now
Chat
This chat starts with Qwen3.8 Flash, not Indra OCR.
≈ 1 credit per reply · See all models
New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.
Start chatting — 5 free credits- Maker
- Datalab
- Type
- Text
- Released
- 2026-10-13
- OpenRouter id
- None
Checked
Indra OCR: facts, with sources
- Datalab’s launch post is dated October 13, 2026 and opens with “Today we’re releasing Indra OCR”, an OCR package built around a 1M-parameter recognition model. The post is by Vik Paruchuri and is filed under product updates. Its first speed claim was 49 pages per second on 32 CPU cores, which the later model card replaces with newer figures (see the table below).Source: datalab.to · checked 2026-11-19
- A second Datalab post, dated October 19, 2026, introduces Indra OCR 1.1. It says the OmniParseBench headline rose from 62.3 to 66.7 (and from 67.0 to 71.4 with math on), handwriting tests from 8.3% to 25.2%, and rotated-page tests from 25.3 to 50.3. The model stays at 1M parameters for the recognizer and about 7M for the whole default pipeline, and still runs on a CPU.Source: datalab.to · checked 2026-11-19
- Datalab’s GitHub releases page lists Indra OCR 1.1.2 as the latest release, published on October 20, 2026. It says the models are the same as in 1.1.0 and that the fix is for digital PDFs whose text layer uses legacy Hindi fonts: Indra now compares the layer’s script with the page image and runs OCR when they disagree, raising word accuracy on such pages from 0.01–0.20 to 0.85–0.89, for about 5 ms more per page.Source: github.com · checked 2026-11-19
- Hugging Face’s API shows the datalab-to/indra-ocr repository was created on October 10, 2026 and last modified on October 19, 2026. These are repository dates, three days before and the same day as Datalab’s two posts.Source: huggingface.co · checked 2026-11-19
- The model card calls Indra OCR a 1M-parameter OCR model that runs on CPU. The default model files total about 76 MB (30 MB of networks and 46 MB of language models for unsure lines); the main Latin recognizer has about 1.25M parameters and all default models together 7.3M, with optional math models extra. It runs on ONNX Runtime with OpenCV and pypdfium2, needs Python 3.10 or newer, no PyTorch and no GPU, and was tested on Linux (x86_64 and aarch64) and macOS on Apple silicon.Source: huggingface.co · checked 2026-11-19
- Input is a PDF or an image file (PNG, JPEG, TIFF, BMP or WebP; one page per TIFF frame). Output is markdown (headings, lists, tables as HTML), plain text, or JSON with a box, text and confidence for every line plus layout blocks in reading order. The block labels are Text, Section-Header, Page-Header, Page-Footer, List-Group, Table, Caption, Footnote, Table-Of-Contents and Form. For digital PDFs it reads the text layer when the layer passes its checks and OCRs only what is missing.Source: huggingface.co · checked 2026-11-19
- Throughput on an AMD EPYC Genoa x86 machine without a GPU, one single-threaded worker per core: Indra OCR reads 6.2 pages per second on 4 cores, 11.9 on 8, 23.7 on 16 and 43.9 on 32, against 4.28 for PaddleOCR’s PP-OCR text models and 2.92 for Tesseract at 32 cores. Datalab adds that throughput depends on the CPU: on other x86 container types it measured 25 to 51 pages per second at 32 cores.Source: huggingface.co · checked 2026-11-19
- Other figures on the card: on OmniDocBench v1.5 (OCR only, 1,651 pages, Datalab’s run without the formula metric) the text edit distance is 0.114 on English pages and 0.237 on simplified Chinese pages (lower is better) and table TEDS is 0.47. On the olmOCR-bench splits it scores 91.2 on headers and footers, 70.0 on multi-column pages, 67.9 on long tiny text, 60.5 on table tests and 20.0 on old scans.Source: huggingface.co · checked 2026-11-19
- Datalab lists its own limits. Complex tables (merged cells, wrapped cell text, nested headers, no ruling lines) can put cells in the wrong rows. Handwriting reads “but not well”: 25.2% on the handwriting tests. Degraded scans score 29.5%. Scripts outside the OCR list such as Syriac, Tibetan and Mongolian only work through a digital PDF’s text layer; written Cantonese characters are not in the Chinese recognizer. Math comes out as plain text unless --math is on, banner-shaped pages read poorly, and pages above 100 megapixels at 200 dpi are rendered at a lower dpi.Source: huggingface.co · checked 2026-11-19
- The code is Apache 2.0; the model weights use a modified OpenRAIL-M license. The card summarises it as free for research, personal use and startups under $20M funding or revenue, and not to be used competitively with Datalab’s API. The license file’s use restrictions bar use for any purpose by someone whose organisation had more than $20,000,000 gross revenue in the prior year or has raised more than $20,000,000 in funding, or that offers a product competing with Datalab’s, except personal or research use; commercial and broader licenses are sold through Datalab. Hugging Face’s metadata tags it “openrail”.Source: huggingface.co · checked 2026-11-19
- The weights are free to download from Hugging Face. To run it: pip install indra-ocr, then indra-ocr paper.pdf for markdown on stdout, --out for one .md per input, --format json or text, --mode auto, ocr or text, --cores to limit processes and --math for LaTeX. The first run downloads the models (about 76 MB). A Python API (indra_ocr.load, conv.convert, convert_many) and a local web viewer (indra-ocr-viewer, port 8765) are included; files stay on your machine.Source: huggingface.co · checked 2026-11-19
- The model card points to Datalab’s managed platform (which runs Datalab’s larger Chandra model) for the best accuracy on handwriting, math, complex tables and degraded scans, and offers $5 in free credits; the GitHub page says about 5,000 free convert pages a month with a work email. Datalab’s pricing page lists Free at $0 with up to $20 of free credit once, Launch at $25 a month and Team at $400 a month, all at per-processor rates. We did not find a per-page price for Indra OCR itself, because the weights are the free part.Source: datalab.to · checked 2026-11-19
- Hugging Face’s API on November 19, 2026 lists the repository with 6,800 downloads, 538 likes and five Spaces, including Datalab’s own and a WebAssembly build. These are Hugging Face counters, not Datalab statements, and they change daily.Source: huggingface.co · checked 2026-11-19 · second-hand
- Indra OCR is not on OpenRouter’s model list (checked November 19, 2026; the list has no Datalab model), so this site cannot run it. This site is a text chat with other models, such as Qwen3.8 Flash and Claude Haiku 5.5.Source: openrouter.ai · checked 2026-11-19 · second-hand
Our table
Indra OCR next to PaddleOCR and Tesseract
- On OmniParseBench (16,288 pass/fail tests on 2,937 pages, 4 vCPUs per engine, one page at a time) the model card gives Indra OCR in auto mode a headline of 66.7, text 46.7, tables 61.1, layout 92.2 and a median of 212 ms per page; with --math 71.4, 62.7, 59.6, 92.0 and 357 ms. PaddleOCR PP-StructureV3 scores 58.4, 55.6, 47.8 and 71.7 at a median of 40,958 ms; Tesseract 5 scores 19.3 on text. These are Datalab’s own figures on its own benchmark, and we did not run them.Source: huggingface.co · checked 2026-11-19
Maker’s benchmarks
Indra OCR scores and speed against PaddleOCR and Tesseract
People who search for Indra OCR mostly want to know one thing: is a 1M-parameter model on a CPU good enough to replace Tesseract or PaddleOCR? By Datalab’s numbers it is far faster than both and ahead on the combined score, but not on everything: PaddleOCR’s full PP-StructureV3 pipeline reads plain text better than Indra OCR in its default mode (55.6 against 46.7) and only falls behind once the --math option is on. All of it comes from Datalab’s own benchmark, not an independent test. This site cannot run an OCR engine, but if you want a model to talk through a document, see Qwen3.8 and Claude Haiku 5.5.
| Measure | Indra OCR (default) | Indra OCR with --math | PaddleOCR PP-StructureV3 | PaddleOCR PP-OCRv6 small (text only) | Tesseract 5 |
|---|---|---|---|---|---|
| OmniParseBench headline | 66.7 | 71.4 | 58.4 | No headline score | No headline score |
| Text | 46.7 | 62.7 | 55.6 | 35.8 | 19.3 |
| Tables | 61.1 | 59.6 | 47.8 | 0.0 | 0.0 |
| Layout | 92.2 | 92.0 | 71.7 | Not scored | Not scored |
| Median milliseconds per page | 212 | 357 | 40,958 | 2,685 | 5,496 |
| Pages per second at 32 cores | 43.9 | About 38% lower than the default | At most 0.18 | 4.28 | 2.92 |
Our table
Indra OCR scores by script on OmniParseBench
- Datalab says it measured 95 languages on held-out printed pages: 83 read at character F1 0.90 or higher and 67 at 0.95 or higher. Four recognizers cover Latin, Cyrillic and Greek; Chinese, Japanese and Korean; Arabic, Persian, Urdu, Pashto, Kurdish and Hebrew; and the Indic, Southeast Asian, Georgian, Armenian and Ethiopic scripts, and a line classifier sends each line to one of them. In text-layer mode on 1,125 digital-PDF pages from 75 language families, it kept at least 98% of the layer’s characters in every family.Source: huggingface.co · checked 2026-11-19
Maker’s benchmarks
Indra OCR OmniParseBench score for selected scripts and the number of tests behind each
The card warns that most scripts have few tests, so one test moves a score by 2 to 8 points. Read the small-sample rows (Arabic with 13 tests, Japanese kana with 19) as hints, not rankings. For Chinese, the card also compares 64.0 for Indra OCR with 65.4 for PaddleOCR’s PP-OCRv5 mobile Chinese text models and 68.4 for PP-StructureV3 on the same 136 tests.
| Script | Score | Tests |
|---|---|---|
| Latin | 61.6 | 9,406 |
| Chinese | 64.0 | 136 |
| Korean | 79.2 | 24 |
| Japanese kana | 57.9 | 19 |
| Arabic | 69.2 | 13 |
| Hebrew | 80.0 | 30 |
| Devanagari | 69.6 | 23 |
| Thai | 87.2 | 39 |
Get told when it is available
Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.
Nearest available
Alternatives to Indra OCR you can use now
Qwen3.8 Flash
Alibaba$0.15 / $0.47 (per million tokens)
- Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
- Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13
Claude Haiku 5.5
Anthropic$0.10 / $0.50 (per million tokens)
- Anthropic’s model page says Claude Haiku 5.5 was released on October 7, 2026, is the latest Haiku, and will not be retired sooner than October 7, 2027. Anthropic’s launch post is dated the same day.Source: platform.claude.com · checked 2026-11-17
- In its launch post Anthropic calls Haiku 5.5 "the cheapest, fastest, and most capable small model we’ve ever released", for high-volume, cost-sensitive tasks such as summaries, compactions, database queries and classification, and as a coding subagent alongside Opus 5.5 and Sonnet 5.5. It calls Haiku 5.5 its first Haiku-class model with an adjustable effort setting.Source: anthropic.com · checked 2026-11-17
- Anthropic’s model page says it is built for high-volume, latency-sensitive work such as classification, routing, extraction and subagent tasks, and rates its latency "Fastest" in the current lineup (Fable 5.2 and Opus 5.5 are "Moderate", Sonnet 5.5 is "Fast").Source: platform.claude.com · checked 2026-11-17
FAQ
Indra OCR FAQ
What is Indra OCR?
Indra OCR is a small OCR package from Datalab that reads PDFs and images and returns markdown, plain text or JSON with a box for every line and a label for every layout block. Its recognition model has about 1M parameters and runs on ONNX Runtime, so it needs no GPU and no PyTorch. The card says it is named for Indra.
How fast and how accurate is Indra OCR?
By Datalab’s own figures, it reads 43.9 pages per second on 32 CPU cores and about 210 ms per page on 4 cores, with an OmniParseBench headline of 66.7 (71.4 with the --math option). PaddleOCR’s full PP-StructureV3 scores 58.4 and takes a median of 40,958 ms per page. These are the maker’s figures on its own benchmark, not an independent test.
How do I run Indra OCR?
Install it with pip install indra-ocr, then run indra-ocr paper.pdf to get markdown, or add --format json for boxes. Python 3.10 or newer is needed, and the first run downloads about 76 MB of models from Hugging Face. There is also a Python API and a local web viewer, so your files stay on your machine.
Which languages does Indra OCR read, and what about handwriting?
Datalab says it reads 80+ languages on scans, from Latin and Cyrillic scripts to Chinese, Japanese, Korean, Arabic, Hebrew, Thai and the Indic scripts. Handwriting is weak by its own numbers: 25.2% on the handwriting tests, and Datalab itself points users to its larger Chandra model for handwriting, math and complex tables. Chinese scores 64.0 on 136 tests.
Is Indra OCR free for commercial use?
Only within limits. The code is Apache 2.0, but the weights use a modified OpenRAIL-M license. The license file bars use by organisations with more than $20,000,000 in prior-year revenue or in total funding raised, and by anyone offering a product that competes with Datalab’s, except for personal or research use. Read the LICENSE file and Datalab’s pricing page before building a product on it.
When was Indra OCR released, and what is 1.1?
Datalab’s first post is dated October 13, 2026. Indra OCR 1.1 followed on October 19, 2026 and added handwriting, 80+ languages on scans and pages at any angle; Datalab says the headline rose from 62.3 to 66.7. The latest release is 1.1.2, from October 20, 2026, which fixes digital PDFs with legacy Hindi fonts.
Can I use Indra OCR on this site?
No. Indra OCR is not on OpenRouter’s model list, and this site is a text chat with other models, so it cannot run an OCR engine over your files. If you want a model to discuss a document, on this site a typical reply to Qwen3.8 Flash costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Hy4 Flash is another small chat model here.
Is this the official Datalab website?
No. This page is unofficial and not affiliated with or endorsed by Datalab. Facts link to the pages we read.