Engine Research — Docling & MarkItDown (verified 2026-07-29)

Docling (docling-project/docling)

• 2.116.0 (2026-07-29), releases every 3–7 days; Python ≥3.10. Hosted by the LF AI & Data Foundation (donated by IBM Research Zurich, still primary maintainers); ~64k stars. Companions: docling-core, docling-serve, docling-jobkit, docling-parse (v7 as of docling 2.106.0).
• Capabilities: PDF layout analysis, reading-order detection, TableFormer table structure, formula/code recognition, picture classification, chart-to-table, OCR via EasyOCR/RapidOCR/Tesseract/VLM (Nanonets OCR2).
• Formats: in — PDF, DOCX, PPTX, XLSX, HTML, EPUB, images, audio (ASR), WebVTT, EML/MSG, LaTeX, ODF, XBRL, video, BoxNote. Out — DoclingDocument; Markdown, HTML, lossless JSON, DocTags, DocLang, WebVTT.
• Production: docling-serve REST (POST /v1/convert/source, async + task polling/WebSocket); container images base/-cpu/-cu128/-cu130 (4.4–11.4 GB); horizontal scaling via Redis RQ workers or Kubeflow; model prefetch with docling-tools models download → artifacts_path for air-gapped deploys. GPU strongly recommended at volume; reuse a loaded DocumentConverter for batch.
• VLM pipeline: granite-docling-258M (superseded SmolDocling); Transformers, MLX (~6 s/page vs ~102 s CPU on Apple silicon), or remote OpenAI-compatible endpoints (vLLM/Ollama preset since 2.66.0).
• Weaknesses: memory — OOM on very large PDFs (killed ~page 345 on 32 GB, docling#3345; 900+ pages, #1654; 3× memory regression, #2786; cross-conversion leak, #2788). Workarounds: pypdfium2 backend, 50–100-page page_range chunks, disable OCR when not needed. CPU-only is seconds-per-page.
• From Node/TS: REST via docling-serve (recommended); official docling-ts / @docling/docling-core npm packages type the output (not conversion); community docling-sdk wraps API + CLI subprocess.

Sources: https://pypi.org/project/docling/ · https://github.com/docling-project/docling/blob/main/CHANGELOG.md · https://github.com/docling-project/docling-serve · https://docling-project.github.io/docling/usage/advanced_options/ · https://docling-project.github.io/docling/usage/vision_models/ · https://github.com/docling-project/docling-ts

MarkItDown (microsoft/markitdown)

• v0.1.7 (2026-07-29; 0.1.6 May 2026); Python ≥3.10; still 0.x (API may move). 0.1.7 fixed O(n²) PPTX chart conversion and LaTeX/OMML math bugs.
• Formats: PDF, DOCX, PPTX, XLSX/XLS, Outlook .msg, images (EXIF+OCR), audio (transcription), HTML, CSV/JSON/XML, ZIP (recursive), YouTube URLs, EPub. Extras: [all], [pdf] (pdfminer-six + pdfplumber), [docx], [pptx], [xlsx], [outlook], [audio-transcription], [az-doc-intel], …
• Philosophy: Markdown as token-efficient LLM-native output; structure over visual fidelity; ~80 MB footprint, no GPU, fast by design (~0.6 s for a 14-page PDF in published comparisons).
• Plugins & MCP: third-party plugins off by default (--use-plugins, #markitdown-plugin tag); official markitdown-mcp (0.0.1a4) — STDIO / Streamable HTTP / SSE, single convert_to_markdown(uri) tool, no auth — localhost only.
• LLM features: pass llm_client + llm_model for image descriptions (standalone images and PPTX-embedded).
• Weaknesses: scanned PDFs → near-empty output (no built-in OCR); multi-column PDFs interleave; complex/merged tables still weak (pdfplumber addition helps simple ones); font-based formatting cues lost in PDFs.
• From Node/TS: no official JS port (discussion #190) — managed subprocess of the real Python package is the most faithful; unofficial markitdown-ts/-js are feature-incomplete.

Sources: https://github.com/microsoft/markitdown · https://pypi.org/project/markitdown/ · https://github.com/microsoft/markitdown/tree/main/packages/markitdown-mcp · https://github.com/microsoft/markitdown/discussions/190

Head-to-head (published third-party numbers)

Metric	MarkItDown	Docling
14-page PDF	~0.6 s	~41 s CPU / ~9 s MLX
Install footprint	~80 MB	~2.4 GB (with VLM)
Complex-table cell accuracy	weak (flattens)	~97.9% (Procycons)
Aggregate quality (one comparison)	~82% F1	~88% F1 (secondhand, treat as directional)
Scanned docs	near-empty	auto-OCR

Sources: https://www.danilchenko.dev/posts/markitdown-vs-docling-vs-marker/ · https://procycons.com/en/blogs/pdf-data-extraction-benchmark/