10 Best Open Source OCR Tools in 2026

For most of the last decade, the answer to "which open source OCR should I use" was Tesseract, and the rest of the list was footnotes. That stopped being true in 2025.

On the official OmniDocBench v1.6 leaderboard, four models now score above 96%: OXR-1.0 at 97.03, TeleOCR at 96.91, OvisOCR2 at 96.47 and PaddleOCR-VL 1.6 at 96.34. None of them is larger than 1.3 billion parameters. Qwen3-VL, a general purpose vision language model with 235 billion parameters, scores 89.78% on the same benchmark. The top score is not the whole story, though: OXR-1.0 needs a paid licence above $2 million in revenue, and the cost of running these models varies by a factor of 30.

That inversion is the story of open source OCR in 2026. Small models trained specifically on documents now beat enormous general models, and they beat the classical engines by a wider margin still. A list written eighteen months ago is not slightly out of date. It is pointing at the wrong tier of the market.

This comparison covers ten engines that are open source, actively maintained, and usable in production. Updated in November 2026: TeleOCR, OvisOCR2 and LightOnOCR-3 join the list, GOT-OCR 2.0 and EasyOCR leave it, and the reasons are explained at the end. If you are new to the field, our guide to what OCR is and what it is used for covers the fundamentals first.

The 10 best open source OCR engines compared

Scored on the official OmniDocBench v1.6 leaderboard where the model is listed (November 2026), with the practical constraints that usually decide the choice: whether you need a GPU, how the engine handles tables and handwriting, and what the licence permits.

1. TeleOCR, the best permissive score

TeleOCR scores 96.91 on the official OmniDocBench v1.6 leaderboard, the best result of any model released under a plain Apache 2.0 licence. It is a 1.2 billion parameter model built on Qwen2.5-VL, published in August 2026 under the name NaviDC-OCR and renamed TeleOCR in September.

The score is the reason to test it first. The caveat is age: the model is three months old, its authors publish no minimum GPU, no table or handwriting breakdown and no olmOCR-Bench result, and there is little public experience of running it in production yet.

Best for: Teams that want the highest score they can use commercially without conditions, and are ready to run their own evaluation.

2. OvisOCR2, the cheapest top model to run

OvisOCR2 scores 96.47 on the OmniDocBench v1.6 leaderboard at 0.8 billion parameters, under Apache 2.0. Its technical report was published in July 2026.

What sets it apart is cost. An independent test on a single consumer RTX 5090 card measured around 560 to 590 pages per minute, which puts the cost at roughly $20 to $28 per million pages on rented hardware. That is the lowest measured figure we found for a model at this accuracy.

Best for: High volume scanned documents where you want top tier accuracy on modest hardware.

3. PaddleOCR-VL 1.6, the mature all rounder

PaddleOCR-VL 1.6 scores 96.34 on OmniDocBench v1.6 at 0.9 billion parameters under Apache 2.0. It is no longer the top score, but it is the most mature option near the top: a large ecosystem, documented deployment, and regular releases from Baidu.

It is strong on text, formulas and tables at the same time, and it holds up on skewed scans, warped pages and screen photos. Language coverage runs past 100, and charts come out as structured HTML.

The trade off is the PaddlePaddle ecosystem: if your stack is PyTorch native, expect some friction. An independent benchmark measured 2.2 pages per second for the first version on an H100, about $476 per million pages.

Best for: Teams that want a top tier model with the most documentation and community behind it.

4. MinerU 4, document parsing as a pipeline

MinerU 4, released in September 2026, pairs a 1.2 billion parameter vision model (95.75 on OmniDocBench v1.6) with a full parsing pipeline: it segments the PDF, orders the reading flow and returns structured output. Version 4 introduced four quality tiers, and the basic one runs on CPU.

For papers, reports and multi column layouts, that pipeline saves real engineering time. The licence has changed: MinerU is no longer AGPL. It now uses its own licence based on Apache 2.0, with a paid licence above 100 million monthly users or $20 million in monthly revenue, and an obligation to say you use MinerU in an online service.

Best for: Document heavy pipelines where layout complexity is the main challenge, including teams that need a CPU option.

5. GLM-OCR, the speed to accuracy pick

GLM-OCR scores 95.22 on OmniDocBench v1.6 at 0.9 billion parameters. Its weights are under MIT and its code under Apache 2.0.

It wins on the combination of accuracy and speed. A third party production run of 116,000 pages on L40S cards reached about 20 pages per second, for roughly $40 per million pages. If you can only test one model first, it is a defensible starting point.

Best for: A first candidate in any evaluation, and production workloads where throughput matters.

6. DeepSeek-OCR 2, built for long documents

DeepSeek-OCR 2, released in January 2026, scores 90.25 on OmniDocBench v1.6. It is a 3 billion parameter model under Apache 2.0; the first DeepSeek-OCR was under MIT.

Its idea is compression: it turns a page into far fewer visual tokens than other models, which keeps long documents cheap to process. The speed figures usually quoted are for version 1: 4.65 pages per second on an H100 in an independent test, about $225 per million pages, and 200,000 pages a day on one A100 according to its authors.

Best for: High volume batch processing where cost per page matters more than the last points of accuracy.

7. olmOCR 2, built for English at scale

olmOCR 2 comes from the Allen Institute for AI under Apache 2.0. The 7 billion parameter model scores 85.74 on the OmniDocBench v1.6 leaderboard and 82.4 on olmOCR-Bench, the benchmark its authors built for long English PDFs.

Its training uses unit tests and reinforcement learning against verifiable outputs. It needs a GPU with at least 12 GB of memory, and its authors put the cost under $200 per million pages. No new model has been released since October 2025.

Best for: Large English corpora and research pipelines where an open training story matters.

8. LightOnOCR-3, small models with strong parsing scores

LightOnOCR-3, released by LightOn in October 2026, comes in three sizes (0.8, 1 and 4 billion parameters), all under Apache 2.0. Its authors report 86.3 on olmOCR-Bench for the 4B model and 85.5 for the 0.8B one; the family is not yet on the OmniDocBench leaderboard.

The small model is fast: 4.78 pages per second at peak on one H100 according to LightOn. It also publishes a separate score for non English documents, which few vendors do.

Best for: Teams that want a recent, permissively licensed model in several sizes, and multilingual corpora.

9. docTR 1.1, the structured document library

docTR, now at version 1.1 (August 2026), is a library rather than a model. It ships PyTorch and TensorFlow backends, separates detection from recognition so you can swap either, and runs on CPU when required.

For predictable documents such as forms and standard invoices, the two stage design gives you control the end to end models do not: you can tune detection independently of recognition.

Best for: Python teams processing high volumes of consistently formatted documents who want control over each stage.

10. Tesseract 5.5.4, still the only real CPU option

Tesseract, now at version 5.5.4 (October 2026), has been maintained since 2005, supports over 100 languages and remains the default answer in most developer conversations. On clean printed text with a simple layout it is still perfectly adequate, and it is the only engine here that runs well without a GPU. That single fact keeps it relevant regardless of benchmark scores. We examined this in detail in our analysis of whether Tesseract is still the best open source OCR, including a benchmark on real invoices.

Its weaknesses are well documented and have not changed: tables collapse into linear text, poor scans degrade output sharply, handwriting is effectively out of scope, and preprocessing does more of the work than most teams expect.

Best for: Clean printed documents, CPU-only environments, and any project where adding GPU infrastructure is not an option.

What it costs to run open source OCR

The licence is free; the GPU is not. The figures below come from published, reproducible measurements, and they are not strictly comparable: page sets, image resolution, batch size and peak versus sustained speed all differ. Use them for orders of magnitude, then measure on your own documents.

For comparison, hosted APIs charge about $1.50 per 1,000 pages for plain text recognition at Google, AWS and Azure, which is $1,500 per million pages, and $4 to $10 per 1,000 pages once layout is included. Mistral OCR 4.1, which has no open weights, costs $4 per 1,000 pages. Self hosting wins clearly on volume; the line it does not show is the engineering time to run, monitor and update the models.

5 things the benchmark scores do not tell you

OmniDocBench is the most useful public signal in this category and it is still a curated corpus, not your documents. Five differences matter more than the ranking, and knowing which ones apply to you will shorten the shortlist faster than any score.

1. Whether you can run it at all

Eight of the ten engines in the table are vision language models and expect GPU inference, typically 8 to 24 GB of VRAM. Tesseract runs on CPU, docTR runs on CPU more slowly, and MinerU 4 has a basic tier that runs on CPU. If there is no GPU in your budget, the leaderboard is largely irrelevant and your real shortlist has three entries on it.

What to check: The model card on Hugging Face states the minimum VRAM. Test at your actual batch size, because per-page memory scales with concurrency.

2. What happens to tables

The overall OmniDocBench score folds in a table TEDS component, so a strong overall number usually means strong tables, but not always. Tesseract flattens a table into linear text and loses the row and column relationships entirely, which is fatal if your documents are financial. Our guide to table detection and extraction explains what the metric measures and where it misleads.

What to check: Run one of your worst tables through two candidates and compare the structure, not the character accuracy.

3. Whether handwriting is in scope

Every vendor claims handwriting support and the gap between claims and output is wider here than anywhere else in OCR. Neat printing works. Cursive and mixed print-and-handwriting pages remain hard for every open-source option. We tested this directly in our OCR handwriting recognition ranking on real samples.

What to check: Your worst handwritten page, not a clean sample. The difference between engines only shows up on difficult input.

4. What the license actually permits

Apache 2.0 and MIT allow commercial use without obligation, and cover most of this list. The traps are elsewhere. OXR-1.0, the top score on the leaderboard, and Chandra 2 are free only for companies under about $2 million in revenue or funding, and neither may be used to build a competing API. jina-ocr-v1 is under a non commercial licence. HunyuanOCR's licence excludes the European Union, the United Kingdom and South Korea. MinerU dropped AGPL for its own licence with revenue thresholds, and dots.ocr adds its own agreement on top of MIT.

What to check: The LICENSE file on the repository today, not what a comparison article said last quarter. Including this one.

5. What you still have to build afterwards

OCR gives you text, and in the better models layout. It does not tell you which fields to trust, whether the totals reconcile, whether this document is a duplicate of one you processed last week, or whether it has been tampered with. Teams processing invoices at volume discover this after the extraction works and the workflow still does not. Field-level confidence scoring and document fraud detection are separate problems from character recognition.

What to check: Write down what happens to the text after extraction. If the answer involves validation, routing or review, that is a second system you have not costed yet.

Also worth knowing

Several projects appear on most lists of this kind and did not make this one, for reasons worth stating.

OXR-1.0 tops the OmniDocBench v1.6 leaderboard at 97.03 (October 2026), but its licence requires a paid licence above $2 million in revenue or funding and forbids a competing API. Xiaomi-OCR-0 (September 2026, Apache 2.0) reports 96.83, a figure from its own authors that is not yet on the official leaderboard. Chandra 2 scores well on olmOCR-Bench with the same revenue threshold as OXR. dots.mocr, the renamed dots.ocr, adds layout grounding but carries its own licence agreement.

EasyOCR and GOT-OCR 2.0 were in the previous version of this list. EasyOCR has not shipped a release since September 2024 and GOT-OCR 2.0 has not been updated since February 2025, so both gave their places to models released in 2026. EasyOCR is still the quickest thing to install for a prototype.

Surya is strong on tables and mathematical notation, with commercial conditions on its weights. Kraken remains excellent for historical manuscripts. OCRmyPDF is a wrapper that adds a text layer with Tesseract underneath. OCRopus and CuneiForm are effectively dormant.

Two further entries are not open source. Mistral OCR 4.1 is a paid API with no open weights. OCR.Space is a freemium hosted service. Both belong in a comparison of hosted OCR services.

How to choose an open source OCR in 2026

Work through the constraints in order, because each one eliminates more candidates than the benchmark does.

Start with hardware. No GPU means Tesseract, docTR or the basic tier of MinerU 4, and the rest of this article is background reading. A GPU opens the full list.

Then the documents. Table-heavy financial documents point to PaddleOCR-VL or MinerU 4. Clean printed text works on anything. Handwriting means a vision-language model and realistic expectations. Mixed-format inbound piles usually need document classification before OCR rather than after it.

Then the license, and check it on the repository rather than trusting any article. Revenue thresholds (OXR-1.0, Chandra 2, MinerU) and non commercial terms (jina-ocr-v1) in particular catch teams who plan to expose a model through a service.

Then run your own documents. Fifty files including your worst scans, across two or three candidates, will tell you more than the entire OmniDocBench table. The benchmark corpus is not your corpus.

Finally, cost the engineering rather than the license. The model is free. GPU hours, tuning, retraining as your document mix drifts, and the validation layer you will eventually need are not. Compare that honestly against a hosted OCR API before assuming self-hosting is cheaper. For some teams it clearly is. For others the free license turns out to be the smallest line in the budget.

FAQs

What is the best open source OCR in 2026?

On the official OmniDocBench v1.6 leaderboard, OXR-1.0 leads at 97.03, followed by TeleOCR at 96.91, OvisOCR2 at 96.47 and PaddleOCR-VL 1.6 at 96.34, all at or under 1.3 billion parameters. Among models you can use commercially without conditions, TeleOCR has the best score and PaddleOCR-VL the most mature ecosystem. Tesseract remains the only engine here that runs well on CPU, and OvisOCR2 and GLM-OCR are the cheapest to run at scale.

Is Tesseract still worth using in 2026?

Yes, for a specific set of jobs. Tesseract handles clean printed text on simple layouts without a GPU, which no modern vision-language model can match on cost. It struggles with tables, poor scans and handwriting. Our separate analysis of whether Tesseract is still the best open-source OCR covers the tradeoff in detail, including a benchmark we ran on real invoices.

Do open source OCR models need a GPU?

Most of the 2025 and 2026 generation do. TeleOCR, OvisOCR2, PaddleOCR-VL, GLM-OCR, DeepSeek-OCR 2, olmOCR and LightOnOCR-3 are vision language models and expect GPU inference, typically 8 to 24 GB of VRAM depending on the model and batch size. Tesseract runs on CPU, docTR runs on CPU more slowly, and MinerU 4 has a basic tier that runs on CPU. If you have no GPU budget, your practical shortlist is short.

Which open source OCR handles tables best?

PaddleOCR-VL and MinerU lead on table structure recognition, both scoring above 95% overall on OmniDocBench v1.6, which includes a table TEDS component. Tesseract flattens tables into linear text and loses the row and column relationships entirely. If your documents are table-heavy, this single difference usually decides the choice. Our guide to table detection and extraction explains what the metric actually measures.

Can open source OCR read handwriting?

Better than it could two years ago, but unevenly. The vision-language models handle neat handwriting reasonably well; cursive and mixed print-and-handwriting pages remain difficult for every open-source option. Tesseract is effectively unusable for handwriting. We tested several engines on real handwritten samples in our OCR handwriting recognition ranking.

Is open source OCR good enough for invoice processing?

For extracting raw text, yes. For extracting structured fields you can post into an ERP, that is a different question. Open-source OCR gives you text and, in the better models, layout. It does not give you field-level confidence scores, validation against business rules, duplicate detection or fraud checks. Teams processing invoices at volume usually end up building that layer themselves or buying it.

What is the difference between open source OCR and an OCR API?

Open-source OCR is a model you host and run yourself: no per-page cost, full data control, and full responsibility for infrastructure, updates and accuracy tuning. An OCR API is a hosted service you call: predictable per-page pricing, no infrastructure, and less control over where data goes. The real comparison is not licence cost but total cost including engineering time.

Which open source OCR licence allows commercial use?

Apache 2.0 and MIT both permit commercial use, and cover TeleOCR, OvisOCR2, PaddleOCR-VL, GLM-OCR, DeepSeek-OCR 2, olmOCR, LightOnOCR-3, Tesseract and docTR. MinerU now uses its own licence based on Apache 2.0, with revenue thresholds. OXR-1.0 and Chandra 2 are free only under about $2 million in revenue, and jina-ocr-v1 is non commercial. Licences in this space change often, so verify on the repository before committing.

Is there a good open source alternative to ABBYY FineReader?

For text extraction, TeleOCR, PaddleOCR-VL or GLM-OCR will match or beat FineReader's recognition on most document types, at no licence cost. What you do not get is the desktop application, the PDF editing, the support contract or the packaged workflow. The gap is not accuracy any more, it is everything around the engine.

How accurate are the OmniDocBench scores?

They are a reasonable relative signal and a poor absolute one. OmniDocBench measures document parsing across text, formula, table and reading order on a curated corpus. Your documents are not that corpus. A model scoring 96% on the benchmark may do noticeably worse on your specific scan quality, language mix or layout. Run a sample of your own files before deciding.

What happened to OCRopus, Kraken and CuneiForm?

They are still on many listicles and should not be. OCRopus has seen little meaningful development for years and its successor projects have absorbed the work. CuneiForm has not shipped a release in over a decade. Kraken remains genuinely useful but only for historical and manuscript work, which is a research niche rather than a general OCR choice.

Why are Mistral OCR and OCR.Space not on this list?

Because neither is open source. Mistral OCR 4.1 is a paid API with no open weights, and OCR.Space is a freemium hosted service. Both are frequently listed as open-source tools, including in an earlier version of this article, and both are commercial products. They may still be the right choice for your project, but they belong in a different comparison.

Sources

• Official OmniDocBench v1.6 end to end leaderboard (opendatalab/OmniDocBench on GitHub), accessed November 2, 2026.
• Model cards and repositories of each engine on Hugging Face and GitHub (scores, licences, release dates), accessed November 2, 2026.
• E2E Networks, measured throughput and cost per million pages on H100 for OCR models released in 2025.
• Independent throughput measurements: GLM-OCR on L40S (Blue Guardrails, July 2026) and OvisOCR2 on RTX 5090 (July 2026).
• Public price pages of Google Document AI, AWS Textract, Azure Document Intelligence and Mistral OCR 4.1, accessed November 2, 2026.
• Koncile, Tesseract OCR: is it still the best open-source OCR in 2026. Internal benchmark on 14 invoices.