Comparison · Same 38 public documents · Open benchmark
anymd vs Marker
Both tools turn PDFs and office files into Markdown for LLMs. Marker, from Datalab, is a Python library and CLI built around its Surya OCR models, and is at its fastest on a GPU. anymd is one native Rust binary, MIT licensed, that is also an MCP server and needs no GPU. On a 4-CPU machine and the 38 public documents of AgentDocBench, anymd scores 96.4 to Marker's 71.0. Marker converted 30 of the 38; on those 30, anymd took 4.64 s instead of 7,104.5 s and returned 20% fewer output tokens.
Install anymd free See the full benchmark
npx -y @sylphx/anymd setup --client cursor,claude-code,codexAt a glance
| anymd | Marker | |
|---|---|---|
| What it is | One native Rust binary: CLI and MCP server | Python library and CLI (marker_single, marker) |
| Install | npx -y @sylphx/anymd setup, Homebrew, Scoop, Cargo or Docker | pip install marker-pdf (marker-pdf[full] for Office files and EPUB), Python 3.10 or newer and PyTorch; its OCR modes start a local Surya model server |
| Hardware | Any CPU | GPU, CPU or Apple MPS; its most accurate modes use a GPU |
| MCP tools | Built in: read, outline, search, inspect | Not part of the package |
| AgentDocBench overall | 96.4 | 71.0 |
| Documents converted, 4 CPUs | 38/38 | 30/38: 6 timed out after 900 s each, 2 CSV files are not supported |
| Time, the 30 documents both converted | 4.64 s | 7,104.5 s |
| Output tokens, the same 30 | 222,618 | 277,779 |
| LLM option | Not needed | --use_llm with Gemini, Claude, OpenAI-compatible, Ollama and other models |
| Output | Markdown with <!-- page N --> anchors; JSON with geometry through inspect | Markdown, JSON, HTML and chunks |
| Licence | MIT | Code under Apache 2.0; model weights under a modified OpenRAIL-M licence, free for research, personal use and startups under US$5M funding or revenue |
| Price | Free, MIT. Optional anymd Pro, one payment | Free within the model licence; a commercial licence from Datalab beyond it |
Measured on the same 38 documents
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. Scores are 0 to 100; higher is better.
| AgentDocBench | anymd 8.2.0 | Marker 2.0.0 |
|---|---|---|
| Overall | 96.4 | 71.0 |
| Documents converted | 38/38 | 30/38 |
| Sentences kept verbatim | 97.7 | 78.9 |
| Text F1 | 97.2 | 79.9 |
| Reading order | 98.8 | 76.8 |
| Table cells F1 | 92.2 | 60.9 |
| Time, the 30 documents both converted | 4.64 s | 7,104.5 s |
| Output tokens, the 30 documents both converted | 222,618 | 277,779 |
Document by document, anymd scores higher on 32, Marker on 5, and 1 is tied.
A document a tool did not convert scores 0. On the 30 documents both converted, the overall is 95.5 for anymd and 89.3 for Marker, and time and tokens above are measured on those 30.
By category
| Category | Documents | anymd | Marker |
|---|---|---|---|
| Papers (math) | 3 | 96.7 | 56.7 |
| Two-column papers | 3 | 98.1 | 65.0 |
| Reports | 2 | 96.9 | 88.2 |
| Financial tables | 4 | 95.2 | 89.5 |
| Forms | 3 | 98.0 | 61.0 |
| Scanned (OCR) | 3 | 77.9 | 53.3 |
| CJK | 3 | 99.1 | 87.6 |
| Slides | 4 | 96.0 | 91.0 |
| Spreadsheets | 4 | 100.0 | 38.1 |
| Word | 3 | 99.9 | 94.7 |
| EPUB | 3 | 98.9 | 66.4 |
| HTML | 3 | 99.6 | 60.6 |
Where Marker scores higher: twocol-bygpt5 (99.0 vs 96.7); scan-nasa-budget (93.3 vs 71.6); scan-nasa-notes (66.7 vs 62.1); slides-epa-1bp (92.7 vs 91.7); epub-markitdown (99.3 vs 96.8).
Where Marker shines: on two of the three scanned PDFs, Marker's OCR beat anymd (93.3 vs 71.6 and 66.7 vs 62.1). Its third scan timed out, which pulls its scanned-category score down.
How it was measured
Both tools ran on GitHub-hosted Linux x86_64 runners with 4 CPUs: Marker 2.0.0 through marker_single on 2026-09-25, in its default CPU mode with no LLM and image extraction off, and anymd 8.2.0 on 2026-10-01. Time is the wall time of a fresh process per file, start-up and model loading included: one timed run after a warm-up for Marker, and the median of three runs for anymd. Each document had 900 s. Marker's README reports much higher throughput on a GPU and with a long-running server; neither was measured, and neither was Marker 2.2.0, released after the run. Tokens are counted with o200k_base.
The benchmark lives in the anymd repository. The scoring does not know which tool produced an output, and the corpus, ground truth, scoring and raw results are public in bench/. Anyone can rerun it or submit a tool.
Why agent builders pick anymd
No GPU, no model server, no PyTorch
npx -y @sylphx/anymd setup adds anymd to Claude Code, Codex, Cursor, VS Code, Claude Desktop, Windsurf and Gemini CLI at once. Homebrew (brew install sylphxai/tap/anymd), Scoop (scoop install anymd from the SylphxAI/scoop-bucket bucket), Cargo (cargo install anymd) and Docker (ghcr.io/sylphxai/anymd) install the same binary. It runs on a laptop, in CI or in a small container, with nothing to download before it converts a born-digital document.
Every document, in seconds, on a CPU
On the 4-CPU runner anymd converted all 38 documents, including the two CSV files and the six that timed out for Marker. On the 30 that both converted, anymd took 4.64 s in total.
Layouts that usually break stay readable
anymd reads PDF glyph positions, so it keeps columns, tables and forms in order: reading order 98.8 vs 76.8 and table cells F1 92.2 vs 60.9. The largest category gaps are spreadsheets (100.0 vs 38.1), math papers (96.7 vs 56.7), HTML (99.6 vs 60.6) and forms (98.0 vs 61.0).
MIT, for any use
anymd is MIT licensed, with no revenue limit and no separate commercial licence. The optional OCR models it can download on request are Apache 2.0.
When Marker is the better fit
- You have a GPU and want vision-model OCR for degraded scans, handwriting or dense equations.
- You want an LLM in the loop (
--use_llm) to merge tables across pages or pull values out of forms. - You need JSON, HTML or chunked output with layout blocks, or Marker's interactive viewer.
Switch in one line
marker_single report.pdf --output_format markdown --output_dir out/ # Marker
anymd report.pdf -o report.md # anymd, after npm install -g @sylphx/anymdFor MCP, add npx -y @sylphx/anymd to your client, or run npx -y @sylphx/anymd setup to add it to every client. With Docker:
docker run --rm -v "$PWD:/data" ghcr.io/sylphxai/anymd report.pdf > report.mdMore install options are in Getting started.
Check what your agent cites with anymd Pro
Converting a file 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. Prefer a hosted service? Use Sylphx AI parse to convert documents without installing a local runtime, and pay per page.
FAQ
Is anymd a drop-in replacement for Marker? For Markdown output, mostly: anymd report.pdf -o report.md replaces marker_single report.pdf --output_format markdown, and npx -y @sylphx/anymd setup adds anymd to your MCP clients. Code that imports marker cannot switch today, because anymd's Python package is not on PyPI yet.
Is anymd faster than Marker? On a 4-CPU runner, yes: on the 30 documents both converted, anymd took 4.64 seconds and Marker 7,104.5 seconds, each timed as a fresh process per file with start-up and model loading included. Marker on a GPU was not measured.
Why does Marker score 71.0? A document a tool did not convert scores 0. Marker timed out on 6 of the 38 documents after 900 seconds each and does not read the 2 CSV files. On the 30 documents both converted, the overall is 95.5 for anymd and 89.3 for Marker.
Does anymd need a GPU or Python? 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.
Can I use anymd commercially? Yes. anymd is MIT licensed, with no revenue limit. Marker's code is Apache 2.0, and its README says its model weights are free for research, personal use and startups under US$5M funding or revenue, with a commercial licence from Datalab beyond that.
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.
Marker is a Datalab project; anymd is not affiliated with or endorsed by Datalab. Benchmark figures come from the AgentDocBench results committed in the anymd repository. Marker details come from its README.