ArchestackGitHub ↗ (opens in a new tab)

Open-source alternatives to LlamaParse

docling leads with 70, ahead of paddleocr (54). Updated from public GitHub data.

The highest-scoring open-source projects in Data and scraping for AI. They are picks, not exact replacements, so read the weaknesses before you switch. The same picks fit if you are leaving Reducto or Azure Document Intelligence.

Most stars
paddleocr 91,043 stars
Try before you install
paddleocr live demo
Open-source alternatives to LlamaParse
RankProjectScore
1doclingPython library and CLI that converts PDFs, Office files and more into structured textData and scraping for AI68.8k starsMIT70 out of 100
2paddleocrOCR and document parsing toolkit that outputs Markdown and JSON from PDFs and images Live demo ↗ (opens in a new tab)Data and scraping for AI91k starsApache-2.054 out of 100

Reviews

170 out of 100

docling

Python library and CLI that converts PDFs, Office files and more into structured text

68.8k starsMITlast commit Oct 2026

Docling parses PDF, DOCX, PPTX, XLSX, HTML, EPUB, images, audio and other formats into a unified DoclingDocument, with PDF layout, reading order, table structure, code and formula recognition. It exports to Markdown, HTML, DocLang, DocTags and lossless JSON. It runs locally via a CLI or Python API, with OCR, VLM support (GraniteDocling), an MCP server and a docling-serve API server.

Strengths

  • Wide input coverage: PDF, Office, ODF, EPUB, email, Apple iWork, XBRL, audio, video
  • Local execution suits sensitive data and air-gapped environments
  • Integrations with LangChain, LlamaIndex, Crew AI and Haystack
  • Ships a CLI, Python API, MCP server and docling-serve HTTP server

Weaknesses

  • Hardware requirements (RAM, GPU) are not stated in the README
  • Individual models carry their own licenses, separate from the MIT codebase
  • Metadata extraction and chemistry understanding are listed as coming soon
  • Requires Python 3.10 or higher; 3.9 support was dropped in 2.70.0
  • GPU optional
  • Docker
  • Models: GraniteDocling, VLMs, OCR engines, ASR models
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
254 out of 100

paddleocr

OCR and document parsing toolkit that outputs Markdown and JSON from PDFs and images

Live demo ↗ (opens in a new tab)91k starsApache-2.0last commit Sep 2026

PaddleOCR runs text recognition (PP-OCRv6) and document parsing (PaddleOCR-VL 0.9B VLM, PP-StructureV3) on PDFs and images, producing Markdown or JSON for RAG pipelines. It runs on CPU, NVIDIA GPU, Kunlunxin XPU and other accelerators, with ONNX, OpenVINO, TensorRT and Transformers backends. Models are on HuggingFace and ModelScope.

Strengths

  • Three PP-OCRv6 tiers (1.5M, 7.7M, 34.5M params) cover edge to server
  • One PP-OCRv6 model handles 50 languages; PaddleOCR-VL covers 111
  • Outputs Markdown, JSON and DOCX; PP-StructureV3 returns table cell and text coordinates
  • Multiple inference backends: Paddle, Transformers, ONNX Runtime, OpenVINO, TensorRT, plus C++ deployment

Weaknesses

  • Depends on the PaddlePaddle framework; GPU installs need matching CUDA builds
  • Benchmark claims such as 96.3% on OmniDocBench are the authors' own
  • Python support listed as 3.8 to 3.12 only
  • Several pipelines and models, so choosing between PP-OCR, VL and StructureV3 takes reading
  • GPU optional
  • Needs PaddlePaddle
  • Models: PP-OCRv6, PaddleOCR-VL-1.6, PP-StructureV3, HPD-Parsing, ONNX
Live demo ↗ (opens in a new tab)Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)