Skip to content

Comparisons · 7 tools · Same 38 public documents · Open benchmark

Document to Markdown converters, compared ​

Seven converters, the same 38 public documents, the same 4-CPU machine. On AgentDocBench, anymd scores 96.4 and converts all 38 documents in 11.5 s. Every figure on this page is generated from the raw results in the anymd repository, including the categories where another tool wins. Start with the scoreboard, or open the head-to-head page for the tool you use today.

Install anymd free    See the full benchmark

bash
npx -y @sylphx/anymd setup --client cursor,claude-code,codex

Scoreboard ​

AgentDocBench checks what an agent needs from a converter: the words verbatim, in reading order, with tables intact, fast, and in as few tokens as possible.

ToolOverallConvertedTable cells F1TimeOutput tokensHead to head
anymd 8.2.096.438/3892.211.5 s368,659–
Docling 2.130.093.038/3889.92,432.4 s430,148anymd vs Docling
Kreuzberg 4.10.481.738/3838.416.0 s455,689anymd vs Kreuzberg
Unstructured 0.27.881.238/3838.4346.5 s414,333anymd vs Unstructured
MarkItDown 0.1.876.838/3857.275.6 s538,388anymd vs MarkItDown
Marker 2.0.071.030/3860.97,104.5 s277,779anymd vs Marker
pdftotext 24.02.042.223/380.00.90 s201,448anymd vs pdftotext

Scores are 0 to 100; higher is better. Overall is the mean of the 12 category scores, and a document a tool did not convert scores 0. Time and tokens cover the documents each tool converted, so the table below compares them on the same 19 documents.

Speed and tokens on the 19 documents every tool converted ​

ToolTimeOutput tokensMeasured
anymd4.58 s146,6602026-10-01
Docling1,723.3 s162,2902026-09-25
Kreuzberg7.94 s139,4222026-09-25
Unstructured187.4 s145,3602026-09-25
MarkItDown45.4 s194,3822026-09-25
Marker5,256.3 s181,1322026-09-25
pdftotext0.66 s147,1832026-09-25

Every tool ran on GitHub-hosted runners (Linux x86_64, 4 CPUs), timed as a fresh process per file.

Which converter fits ​

Docling is the closest to anymd on accuracy. Pick it for an in-process Python library with LangChain or LlamaIndex integrations, a lossless JSON document model, or vision-language models on a GPU. It scored higher on designed reports, scans and EPUB; on a CPU it was the slowest of the tools that converted every document.

Kreuzberg has a Rust core with bindings for Python, TypeScript, Go, Java and more, reads 91+ formats including email and archives, and adds chunking and embeddings. It scored highest on HTML and tied anymd on spreadsheets; tables are its weak spot.

Unstructured splits files into typed elements, with chunking and connectors for batch RAG ingestion in Python. Its text is close to anymd's; its tables are not.

MarkItDown is Microsoft's lightweight Python converter, good on Office files, spreadsheets, HTML and designed reports. With its default options it returns no text for scanned PDFs.

Marker uses vision-model OCR and an optional LLM, and is at its best on a GPU. Its OCR beat anymd on two of the three scans; on a 4-CPU runner it timed out on six documents.

pdftotext is Poppler's plain-text extractor: the quickest way to pull a born-digital PDF's text layer, often already installed. It reads PDFs only, with no OCR and no tables.

anymd is for agents that read PDFs, Office files, EPUB, HTML, scans and media through one MCP server on any CPU, with page anchors to cite and tables kept as Markdown. It is free and MIT licensed.

How the scores were measured ​

Every tool ran on GitHub-hosted Linux x86_64 runners with 4 CPUs, through its own CLI or Python API with default options unless its page says otherwise. Time is the wall time of a fresh process per file, start-up and model loading included. The scoring does not know which tool produced an output. GPUs, long-running servers and releases newer than the ones listed were not measured. The corpus, ground truth, scoring and raw results are public in bench/, and anyone can rerun them or submit a tool.

Check what your agent cites with anymd Pro ​

Converting a file with anymd is free. When an agent quotes a PDF or a video, anymd Pro takes you back to the source:

  • Cite-check compares up to 100 quotes per request with the page and area each one cites, and reports whether it matches, doesn't match, or needs more evidence. It uses the text and its position on the page, not an AI opinion.
  • Video timelines and frames show scenes, chapters and subtitle or transcript timings, and the frame behind a cited moment with its actual decoded timestamp and a SHA-256 hash of the image.

Pay once, US$29 or the local price in 16 currencies, for a 100-year licence for one person. Pro works offline, with no subscription and no account.

See anymd Pro

FAQ ​

Which document to Markdown converter is the most accurate? On AgentDocBench, anymd scores 96.4 overall across 38 documents in 12 categories, followed by Docling. The scoreboard lists every tool's score, time and output tokens, and each head-to-head page lists the categories and documents where the other tool scores higher.

Which converter is the fastest? pdftotext, on the PDFs it reads, because it copies the text layer and runs no OCR. Of the tools that converted all 38 documents, anymd is the fastest, at 11.5 seconds on a 4-CPU runner.

Does anymd need Python or a GPU? No. anymd is a native Rust binary that runs on any CPU. The npx install needs Node 18 or newer; Homebrew, Scoop and Docker install the same binary without Node.

Is the benchmark biased toward anymd? It lives in the anymd repository, so you can check it: the scoring does not know which tool produced an output, the corpus and ground truth are public, and the categories where anymd loses are published. Anyone can rerun it or add a tool.

Is anymd free? Yes. anymd is MIT licensed, including the CLI, the MCP server, OCR and transcripts. anymd Pro is an optional one-time purchase, US$29 or the local price in 16 currencies, that adds PDF cite-check and video timelines and frames.

Docling, Kreuzberg, Unstructured, MarkItDown, Marker and Poppler are projects of their respective owners; anymd is not affiliated with or endorsed by any of them. Benchmark figures come from the AgentDocBench results committed in the anymd repository.