The open-source OCR landscape has exploded since mid-2026. Between June and August, at least eight new models launched or gained serious traction. The old Tesseract-vs-everything dynamic is gone. You now have real choices across the quality, speed, and language spectrum.
I’ve tested all of these. Some are excellent. Some are overhyped. Here’s the current state.
The Comparison Table (Updated August 2026)
| Model | Params | License | Languages | Multi-page | Tables | Equations | Min Hardware | Best For |
|---|---|---|---|---|---|---|---|---|
| Baidu Unlimited-OCR | 3B | MIT | 40+ | Yes (single pass) | HTML | LaTeX | 8 GB RAM | Multi-page PDFs, structured output |
| Surya | 650M | Apache 2.0 | 90+ | No | Yes | Limited | 4 GB VRAM | Speed, small footprint, multilingual |
| dots.ocr | 1.7B | Apache 2.0 | Any script | No | Yes | LaTeX | 6 GB VRAM | Multilingual document parsing |
| PaddleOCR-VL (PP-OCRv6) | 34.5M | Apache 2.0 | 100+ | No | Yes | No | 2 GB VRAM | Ultra-lightweight, mobile/edge |
| Dolphin | 400M+ | Apache 2.0 | 50+ | No | Yes | Limited | 4 GB VRAM | Complex layouts, analyze-then-parse |
| GLM-OCR | ~500M | Apache 2.0 | 30+ | No | Limited | LaTeX | 4 GB VRAM | Chinese documents, academic content |
| LightOnOCR-2 | 1B | Apache 2.0 | 20+ | No | Limited | Limited | 4 GB VRAM | French/European languages |
| GOT-OCR 2.0 | 580M | Apache 2.0 | 20+ | No | Limited | LaTeX | 4 GB VRAM | Academic papers, sheet music |
| Florence-2 | 770M | MIT | 100+ | No | No | No | 4 GB VRAM | General vision tasks, captioning |
| Nougat | 350M | CC-BY-NC | English-focused | No | Yes | LaTeX | 4 GB VRAM | Academic PDFs, arXiv papers |
| Tesseract 5 | N/A | Apache 2.0 | 100+ | No | No | No | 512 MB RAM | Simple text, minimal hardware |
| DeepSeek-OCR 2 | 1.3B | MIT | 30+ | No | Limited | LaTeX | 6 GB VRAM | Chinese/English documents |
The New Wave: Models Launched June-August 2026
The second half of 2026 brought a flood of new open-source OCR models. Here’s what changed and why it matters.
Surya: The Speed King
650M params, 83.3% on olmOCR-bench, top-scoring model under 3B parameters.
Surya from VikParuchuri (the Tesseract contributor turned independent researcher) focuses on what matters for production: speed, accuracy, and language breadth. At 650M parameters, it runs on modest hardware while outperforming models 5x its size on standard benchmarks.
What makes it different:
- 90+ language support (broadest among neural models)
- Table and layout recognition built-in
- Fast inference even on CPU (optimized architecture)
- Active development with frequent updates
Best for: Teams that need fast, accurate OCR across many languages without massive GPU infrastructure. If you process documents in 10+ languages regularly, Surya is the best open option.
See our Surya complete guide for setup and benchmarks.
dots.ocr: The Multilingual Specialist
1.7B params, SOTA multilingual document parsing, recognizes virtually any script.
dots.ocr from dots.llm takes a different approach: it’s trained to recognize any script or writing system, not just the common ones. If you process documents in less-common scripts (Thai, Arabic, Devanagari, Cyrillic), dots.ocr handles them better than models trained primarily on Latin and CJK.
What makes it different:
- Script-agnostic: works on virtually any writing system
- Strong table extraction with cell-level bounding boxes
- LaTeX equation support
- Good at mixed-language documents
Best for: Multilingual document processing, especially in regions with diverse scripts (India, Southeast Asia, Middle East). Also strong for academic content with mixed languages and equations.
See our dots.ocr complete guide for setup and benchmarks.
PaddleOCR-VL / PP-OCRv6: The Tiny Giant
34.5M params, beats Qwen3-VL-235B and GPT-5.5 on document understanding.
This is from Baidu’s own PaddlePaddle team, and it’s genuinely impressive. At 34.5 million parameters (not billion), it achieves document understanding scores that rival models 1000x its size. The trick: it’s a specialized vision-language model, not a general-purpose one.
What makes it different:
- Incredibly small: 34.5M params runs on phones and edge devices
- 100+ language support
- Apache 2.0 license
- Built on PaddlePaddle (Baidu’s own framework)
- Beats its own sibling (Baidu Unlimited-OCR) on some benchmarks
The comparison hook: Baidu has two OCR models now. Unlimited-OCR (3B, MIT) for maximum quality and PaddleOCR-VL (34.5M, Apache 2.0) for maximum efficiency. Choosing between them depends on your hardware constraints and accuracy requirements.
See our PaddleOCR-VL complete guide for setup and the full comparison.
Dolphin (ByteDance): The Layout Expert
Analyze-then-parse approach for complex document layouts.
Dolphin from ByteDance uses a two-stage approach: first analyze the document structure (identify tables, figures, text blocks, headers), then parse each element with specialized extractors. This gives it an edge on complex layouts where other models struggle.
What makes it different:
- Two-stage pipeline handles complex layouts better
- Good at distinguishing content types (text vs table vs figure)
- Handles multi-column documents well
- Active development from ByteDance
Best for: Documents with complex layouts: multi-column pages, mixed text-and-table pages, forms with embedded tables. If your documents have tricky layouts that trip up single-pass models, Dolphin is worth trying.
See our Dolphin OCR complete guide for setup and benchmarks.
GLM-OCR: The Academic Specialist
GLM-V encoder-decoder with Multi-Token Prediction loss.
GLM-OCR from Zhipu AI (the team behind GLM-4) brings the GLM architecture to document understanding. It uses Multi-Token Prediction loss during training, which helps it handle rare characters and domain-specific terminology better than standard OCR models.
What makes it different:
- Strong on Chinese academic content
- Multi-Token Prediction helps with rare/technical characters
- LaTeX equation support
- Good integration with the GLM ecosystem
Best for: Chinese academic documents, technical papers with specialized terminology, content that mixes Chinese and English. If you’re already in the GLM ecosystem, GLM-OCR is a natural fit.
See our GLM-OCR complete guide for setup and benchmarks.
LightOnOCR-2: The European Contender
1B params, claims to beat DeepSeek-OCR, unverified at this size.
LightOnOCR-2 from LightOn (a French AI lab) focuses on European language performance. While most OCR models optimize for English and Chinese first, LightOnOCR-2 gives strong performance on French, German, Spanish, and other European languages.
What makes it different:
- Strong European language support
- Competitive benchmarks (though self-reported)
- Active development from a European team
- Apache 2.0 license
Best for: European document processing, especially in regulated industries where data sovereignty matters. If you process French, German, or Spanish documents primarily, LightOnOCR-2 is worth benchmarking against the alternatives.
See our LightOnOCR-2 complete guide for setup and benchmarks.
The Existing Models: What’s Still Relevant
Baidu Unlimited-OCR
Still the best all-rounder if you need multi-page PDF support and structured output. At 3B params with MIT license, it’s the most capable open model for general document processing. But the gap is closing. For single-page work, Surya and PaddleOCR-VL are competitive or better.
See our Baidu Unlimited-OCR complete guide for setup.
GOT-OCR 2.0
Still the best for academic papers with heavy equations and sheet music. The 580M param size is efficient. But dots.ocr and GLM-OCR are catching up on academic content while offering better multilingual support.
Florence-2
Still useful as a general vision model that happens to do OCR. If you’re already using Florence-2 for other tasks, adding OCR is free. But dedicated OCR models now significantly outperform it on document-specific benchmarks.
Nougat
Still excellent for arXiv-style academic papers. But the CC-BY-NC license limits commercial use, and there’s been no active development since launch. For commercial academic processing, GOT-OCR 2.0 or dots.ocr are better choices.
Tesseract
Still the only option for no-GPU environments. Still terrible at anything beyond simple text. Use it as a pre-filter or fallback, not as your primary OCR engine.
How to Choose
Start here:
“I need the best open-source OCR with no restrictions” Baidu Unlimited-OCR for multi-page PDFs. Surya for single-page speed. dots.ocr for multilingual scripts.
“I’m processing academic papers” GOT-OCR 2.0 for equations and sheet music. dots.ocr for multilingual academic content. GLM-OCR for Chinese academic papers.
“I have very limited hardware” PaddleOCR-VL (34.5M params) runs on phones. Tesseract if you have no GPU at all. Surya for a good balance of quality and speed.
“I need 100+ language support” PaddleOCR-VL (100+), Florence-2 (100+), or Surya (90+). For quality with fewer languages, Unlimited-OCR.
“I process multi-page documents” Only Baidu Unlimited-OCR handles this in a single pass. Everything else requires page-by-page processing.
“I process complex layouts” Dolphin for analyze-then-parse approach. Surya for built-in layout recognition.
“I need this for production commercial use” MIT or Apache 2.0 licensed options: Unlimited-OCR, Surya, dots.ocr, PaddleOCR-VL, Dolphin, Florence-2, GOT-OCR, Tesseract. Avoid Nougat (CC-BY-NC).
The Cloud Alternative
Sometimes open-source isn’t the right call. If you need:
- The absolute best accuracy (especially on hard documents)
- Enterprise SLA and support
- Zero infrastructure management
- 170 language support
Then paid services like Mistral OCR 4 ($4/1K pages) or Google Document AI ($5/1K pages) are worth considering. See our multimodal AI APIs price comparison for the full cloud pricing landscape.
The sweet spot for many teams: use open-source models for development, testing, and standard documents, then route the hard cases to a cloud API. This hybrid approach gives you the best of both worlds.
What’s Coming Next
The open-source OCR space is accelerating. Things to watch:
- Larger context windows enabling even longer documents in single passes
- Better handwriting recognition (still a gap for all open models)
- More languages with fewer parameters
- Native video/scanned-book OCR (page turning detection)
- Better integration with RAG pipelines (chunking-aware extraction)
- Edge deployment: PaddleOCR-VL shows that useful OCR can run on phones
2026 has already been the best year for open-source OCR. The gap between free and paid is shrinking fast.
FAQ
Which open-source OCR model has the best accuracy?
For general documents, Baidu Unlimited-OCR leads. For multilingual scripts, dots.ocr. For speed-optimized accuracy, Surya (83.3% olmOCR-bench at 650M params). For academic papers, GOT-OCR 2.0. None match Mistral OCR 4’s 72% blind test win rate, but for most documents the difference is negligible.
Can any of these replace Google Document AI or Mistral OCR 4?
For standard business documents (invoices, contracts, forms): yes, Unlimited-OCR and Surya produce usable output for most cases. For edge cases (low quality scans, rare languages, complex nested tables), cloud services still have an advantage. Many teams use a hybrid approach.
Which model runs fastest on consumer hardware?
PaddleOCR-VL (34.5M params) is the lightest neural option. Tesseract is the lightest overall (no GPU needed). Among quality models, Surya (650M) and GOT-OCR 2.0 (580M) are fastest. Unlimited-OCR (3B) is larger but offers the best quality.
Is there an open-source model that handles handwriting?
None of these handle handwriting well. It’s the biggest remaining gap in open-source OCR. For handwritten documents, cloud services (Google, Mistral) still significantly outperform open alternatives.
Can I fine-tune these models for my specific documents?
Yes, all the MIT/Apache-licensed models can be fine-tuned. Unlimited-OCR, Florence-2, and Surya have the most community resources for fine-tuning. Fine-tuning on your specific document types can dramatically improve accuracy.
Do any of these support real-time OCR (video/camera)?
Not directly. These are designed for static document images. For real-time camera OCR, you’d typically use a lighter detection model to find text regions, then feed cropped regions to one of these models. PaddleOCR-VL at 34.5M params is the closest to real-time capable.