Self-hosted OCR and document AI apps for small businesses
Short answer
We track 9 active self-hostable apps for OCR, PDF parsing and AI document processing: PaddleOCR, MinerU, Paperless-ngx, OCRmyPDF, olmOCR, Unstract, Paperless-AI, paperless-gpt, Docling Serve.
Most popular on GitHub: PaddleOCR (92.2k stars).
Run on an ordinary CPU server according to their own docs: PaddleOCR, MinerU, OCRmyPDF, Docling Serve.
Need a GPU for practical use: olmOCR (Required for local inference: a recent NVIDIA GPU with at least 12 GB of GPU RAM (tested on RTX 4090, L40S, A100, H100). Without a GPU you can point it at a remote vLLM or OpenAI-compatible server.).
Lowest official hardware: MinerU (minimum 2 GB RAM, from €11.99/month on our reference server).
Permissive license (MIT/Apache-2.0, fewest conditions): PaddleOCR, olmOCR, Paperless-AI, paperless-gpt, Docling Serve.
No official CPU/RAM figure published: PaddleOCR, Paperless-ngx, OCRmyPDF, olmOCR, Paperless-AI, paperless-gpt, Docling Serve.
Last updated · generated from the same verified data as each app page
Comparison table
| App | What it is | GPU | Official hardware | Est. server cost | License | GitHub stars |
|---|---|---|---|---|---|---|
| PaddleOCR | OCR toolkit and document AI engine that turns scans, images and PDFs into LLM-ready text, Markdown or JSON. | Optional. The server uses a GPU by default if available, otherwise the CPU. | Not published | Not estimated (no official RAM figure) | Apache-2.0 (Permissive) | 92.2k |
| MinerU | Document parser that converts PDFs, scans and Office files into clean Markdown or JSON for LLMs and RAG. | Optional. The basic tier runs ONNX models on CPU; the standard/advanced tiers work on CPU with 8 GB RAM (Vulkan recommended). PyTorch and vLLM/lmdeploy engines need a GPU with 8 GB+ VRAM and 16 GB RAM. The official Docker image targets NVIDIA GPUs only. | minimum 2 GB RAM | from €11.99/mo | MinerU Open Source License (Apache-2.0 with additional terms) (Custom / conditional) | 83.0k |
| Paperless-ngx | Document management system that OCRs scans into a searchable archive, auto-tags them and adds optional LLM suggestions and chat. | No. OCR and the built-in classifier run on CPU; LLM features call the provider you configure. | Not published | Not estimated (no official RAM figure) | GPL-3.0 (Copyleft (GPL)) | 47.6k |
| OCRmyPDF | Command-line tool and Docker image that adds a searchable OCR text layer to scanned PDFs, with a watched-folder service. | No. Tesseract runs on CPU; the optional EasyOCR plugin is the one that recommends a GPU. | Not published | Not estimated (no official RAM figure) | MPL-2.0 (Copyleft (GPL)) | 35.2k |
| olmOCR | Ai2's vision-language OCR that turns PDFs and scans into clean Markdown, including tables, equations and handwriting. | Required for local inference: a recent NVIDIA GPU with at least 12 GB of GPU RAM (tested on RTX 4090, L40S, A100, H100). Without a GPU you can point it at a remote vLLM or OpenAI-compatible server. | Not published | GPU server (not estimated) | Apache-2.0 (Permissive) | 19.8k |
| Unstract | No-code LLM platform that turns invoices, forms and other documents into structured JSON. | Not required by the platform itself; LLMs are reached through the provider you configure (cloud API or a local server such as Ollama). | minimum 8 GB RAM | from €35.99/mo | AGPL-3.0 (Network copyleft (AGPL)) | 7.4k |
| Paperless-AI | AI auto-tagging and document chat for Paperless-ngx (currently unmaintained). | Not stated | Not published | Not estimated (no official RAM figure) | MIT (Permissive) | 6.0k |
| paperless-gpt | LLM add-on for paperless-ngx that does better OCR and writes titles, tags and correspondents. | Not needed for the app itself. A GPU only matters if you run the LLM locally (for example with Ollama). | Not published | Not estimated (no official RAM figure) | MIT (Permissive) | 2.8k |
| Docling Serve | Docling document conversion (PDF, Office, images to Markdown/JSON) as a self-hosted API with a demo UI. | CPU-only image (docling-serve-cpu) is published; CUDA 12.8 and 13.0 images are available for NVIDIA GPUs, and an AMD ROCm image can be built locally. DOCLING_DEVICE picks cpu, cuda or mps, otherwise the best device is chosen. | Not published | Not estimated (no official RAM figure) | MIT (Permissive) | 1.9k |
Apps from our "Document processing" category, sorted by GitHub stars; archived projects are excluded. "CPU server" means the project's own docs say the GPU is optional or not needed, or its official requirements list CPU and RAM only; "needs a GPU" means its docs say so for practical use. We do not rank by OCR quality. Cost = cheapest Hetzner Cloud Regular Performance (shared AMD vCPU) plan that meets the project's official RAM and vCPU figure, list price read on 2026-11-01, excl. VAT, server only (no LLM API usage, backups or your time).
Frequently asked questions
What is the best self-hosted OCR app for a small business?
Most popular on GitHub: PaddleOCR (92.2k stars). Run on an ordinary CPU server according to their own docs: PaddleOCR, MinerU, OCRmyPDF, Docling Serve. Need a GPU for practical use: olmOCR (Required for local inference: a recent NVIDIA GPU with at least 12 GB of GPU RAM (tested on RTX 4090, L40S, A100, H100). Without a GPU you can point it at a remote vLLM or OpenAI-compatible server.). Lowest official hardware: MinerU (minimum 2 GB RAM, from €11.99/month on our reference server). Permissive license (MIT/Apache-2.0, fewest conditions): PaddleOCR, olmOCR, Paperless-AI, paperless-gpt, Docling Serve. No official CPU/RAM figure published: PaddleOCR, Paperless-ngx, OCRmyPDF, olmOCR, Paperless-AI, paperless-gpt, Docling Serve. The right choice depends on your documents and whether you have a GPU.
Which self-hosted OCR and document apps run without a GPU?
PaddleOCR: GPU optional (CPU mode documented) MinerU: GPU optional (CPU mode documented) OCRmyPDF: GPU optional (CPU mode documented) Docling Serve: GPU optional (CPU mode documented)
Which document AI apps need a GPU?
olmOCR: Required for local inference: a recent NVIDIA GPU with at least 12 GB of GPU RAM (tested on RTX 4090, L40S, A100, H100). Without a GPU you can point it at a remote vLLM or OpenAI-compatible server.
Which self-hosted document apps have an MIT or Apache license?
PaddleOCR, olmOCR, Paperless-AI, paperless-gpt, Docling Serve. MinerU uses MinerU Open Source License (Apache-2.0 with additional terms). Paperless-ngx uses GPL-3.0. OCRmyPDF uses MPL-2.0. Unstract uses AGPL-3.0.
Related
- Document processing category
- Self-hosted AI apps that run without a GPU
- Self-hosted RAG and knowledge base apps
- Cheapest self-hosted ChatGPT alternatives
- Self-hosted AI apps that run without a GPU
- Self-hosted RAG and knowledge base apps for small businesses
- Self-hosted AI apps by license: what a business may do
- Which self-hosted AI apps include SSO for free?
- Which self-hosted AI apps send telemetry by default, and how to turn it off
- Self-hosted AI agent and chatbot builders for small businesses
Every figure on this page comes from the linked app pages, which cite the project's own repository and docs. Spotted something wrong? Tell us.