An open-source OCR toolkit developed by Ai2 (the Allen Institute for AI). It converts PDFs and scanned images into clean text and Markdown, handling tables, equations, and handwriting along the way. Rather than simply picking out characters like a conventional OCR engine, a vision language model (VLM) reads the layout of the whole page before transcribing it, so it can infer the correct reading order in two-column layouts and automatically strip out non-body elements such as headers and footers. The first version was released in February 2025, and “olmOCR 2,” which significantly improved accuracy, arrived in October 2025. It is licensed under Apache 2.0, with the model weights published as well.
Key Features
- Layout-aware transcription: Even on complex pages with two columns or text wrapping around figures, it reorders the text to follow the sequence a human would read. Page numbers, headers, and footers are automatically excluded from the body text
- Tables, equations, and handwriting: Tables can be emitted as HTML tables and equations as LaTeX, so structure is not lost in conversion. Handwritten text is supported as well
- Built on a vision language model (olmOCR 2): Based on Qwen2.5-VL-7B and trained with reinforcement learning that uses verifiable tests — such as “is the table structure preserved?” and “is the reading order correct?” — as rewards. Its overall score on the olmOCR-Bench benchmark is 82.4, with notable gains on tables (84.9%) and multi-column text (83.7%)
- Open source and self-hostable: Released under Apache 2.0, with the toolkit on GitHub and model weights (BF16 and FP8) on Hugging Face. You can keep everything inside your own GPU environment without sending documents to an outside service
- Low cost at high volume: Ai2 estimates less than USD 200 per million pages when self-hosted. The FP8 quantized version is reported to run at around 3,400 tokens per second on an H100 GPU
- Try it in the browser: A demo of olmOCR 2 is available on the Ai2 Playground (playground.allenai.org), so you can check conversion results without installing anything
Pricing
| Usage type | Cost | Details |
|---|---|---|
| Online demo (Ai2 Playground) | Free | Try conversion results in the browser |
| Self-hosted (your own GPU) | Software is free (Apache 2.0) | GPU costs only. Ai2 estimates under USD 200 per million pages |
| Via external inference providers | Usage-based pricing per provider (check directly) | DeepInfra, Parasail, and Cirrascale are supported. See each provider’s site for rates |
Pricing is current as of August 2026. Please check the official site and the GitHub repository for the latest information.
Pros & Cons
✅ Pros
- Fully open source down to the model weights, with no vendor lock-in to worry about
- Documents can be processed entirely inside your own environment, which makes it easier to handle confidential material and personal information
- Covers tables, equations, and handwriting — the elements traditional OCR has struggled with
- Very low unit cost at scale, making it suitable for digitizing archives of hundreds of thousands of pages
- Markdown output feeds naturally into RAG (retrieval-augmented generation) pipelines and documentation platforms
⚠️ Cons
- Developed and evaluated on English-language documents; accuracy on Japanese documents is not guaranteed
- Requires an NVIDIA GPU (12 GB VRAM or more) and a CUDA environment; it is not practical to run on a CPU-only machine
- Assumes familiarity with the command line and Python, so the setup bar is higher than a no-code commercial OCR service
- Input is limited to PDF, PNG, and JPEG. Office files must be converted to PDF first
- Because it is generative model based, it can occasionally output text that is not in the source, so important documents still need a visual check
Comparison with Similar Services
| Criteria | olmOCR | Mistral OCR | Marker | Google Document AI |
|---|---|---|---|---|
| Provider | Ai2 | Mistral AI | Datalab | Google Cloud |
| License / delivery | Open source (Apache 2.0) | API (cloud) | Open source + commercial license | Cloud service |
| Self-hosting | Yes (GPU required) | No | Yes (GPU recommended) | No |
| Output formats | Markdown / HTML tables / LaTeX | Mainly Markdown | Markdown / JSON / HTML | Structured JSON |
| Multilingual support | English-focused | Multilingual | Multilingual | Multilingual |
| Cost | Software free (GPU costs only) | Usage-based | Free tier + commercial terms | Usage-based |
Who Is It For
- Researchers and engineers who want to turn English papers, technical documents, and reports into text in bulk as material for search or RAG
- IT staff at companies that cannot send confidential documents to an external API and need everything to stay on in-house GPUs
- Libraries and archive operators digitizing hundreds of thousands of pages of paper and PDF material at the lowest possible cost
- Anyone losing time to manual cleanup because their current OCR mangles tables and equations
- Open-source-minded developers who want to inspect and modify how the OCR works
Summary
What sets olmOCR apart is that it publishes “layout-aware OCR” powered by a vision language model as open source, model weights included. It handles tables, equations, and handwriting, keeps the unit cost of bulk processing low, and is an especially natural fit where confidential documents cannot leave the building. On the other hand, it is worth keeping in mind that it is designed primarily around English documents and requires an NVIDIA GPU environment. A realistic path is to try your own documents on the Ai2 Playground first, confirm that the accuracy meets your requirements, and only then move on to building your own setup.