PDF to Markdown vs Marker
Marker is an excellent open-source library you self-host. pdf2md.dev is a hosted service that converts a PDF in seconds with no install and no GPU – in the browser, by REST API, or from a hosted MCP. Here is an honest side-by-side so you can pick the right one.
Sources and a fair comparison
Documentation reviewed 8 November 2026. Source claims describe available workflows, not an independent accuracy ranking. Test your own PDFs before choosing. Vendor pricing, licences and model aliases can change.
Marker code is Apache 2.0. Its model weights have a separate modified AI Pubs Open Rail-M licence: the published terms distinguish research, personal use and startups under $5M funding/revenue from other commercial use. Check the current code and weight licences for your use.
Marker · Docling · MinerU · MarkItDown · LlamaParse · Adobe · Mistral OCR
Check the output, then use it
Documentation review: 8 November 2026. Compare workflow and output on the same PDFs, settings and pinned versions. We have not run a new cross-tool accuracy benchmark. The published speed dataset remains historical, 10–24 June 2026.
Server processing; temporary results. Review numbers and OCR mistakes against the original.
Self-host Marker, or let pdf2md.dev host it for you
Marker supports an optional --use_llm mode and CPU, GPU or MPS execution. Model configuration, hardware and external LLM costs differ. We have not established an accuracy winner on a shared test corpus; compare the same PDFs with pinned versions and settings.
pdf2md.dev vs Marker, feature by feature
Both produce clean Markdown with tables, formulas and OCR. The real difference is hosted-for-you versus self-hosted.
| pdf2md.dev | Marker | |
|---|---|---|
| Setup | Nothing to install – browser, REST API or hosted MCP | pip install marker-pdf + a Python environment |
| Hardware | None – we host it; no GPU | Runs on CPU/MPS, but built for a GPU (fast in batch on an H100) |
| Cost | Free anonymous tier; paid tiers raise limits | Free to self-host under the licence; you pay GPU + ops. Datalab API is paid |
| Account | None to convert | None to self-host; an API key for the Datalab hosted API |
| Scanned PDF OCR | Built in, many languages, no flags | Yes, via Surya OCR when self-hosted |
| Tables & formulas | Yes – real Markdown tables, math kept | Yes – tables, equations and inline math |
| Input formats | PDF and images, by upload or URL | PDF, image, DOCX, PPTX, XLSX, HTML, EPUB |
| Highest-accuracy mode | Pick the engine – MinerU (dense) or Docling (clean, fast) | Marker supports an optional --use_llm mode and CPU, GPU or MPS execution. Model configuration, hardware and external LLM costs differ. We have not established an accuracy winner on a shared test corpus; compare the same PDFs with pinned versions and settings. |
| Automation | REST API + hosted MCP for agents | Python library; optional paid Datalab API |
| Licensing | Hosted SaaS; your files are not used to train models | Marker code is Apache 2.0. Its model weights have a separate modified AI Pubs Open Rail-M licence: the published terms distinguish research, personal use and startups under $5M funding/revenue from other commercial use. Check the current code and weight licences for your use. |
Marker details from its public project documentation; pdf2md.dev limits are the current free-tier values. Both projects evolve – check each source for the latest.
More options? See the full roundup of the best PDF to Markdown converters for the whole field at a glance.
When Marker is the better choice
Marker is a genuinely strong tool. Reach for it when these fit.
Fully self-hosted
You need conversion to run entirely on your own machines, offline or on-prem, with no third-party service in the path.
Many input formats
You convert DOCX, PPTX, XLSX, HTML or EPUB as well as PDF, and want one tool for all of them.
LLM-boosted accuracy
Marker supports an optional --use_llm mode and CPU, GPU or MPS execution. Model configuration, hardware and external LLM costs differ. We have not established an accuracy winner on a shared test corpus; compare the same PDFs with pinned versions and settings.
When pdf2md.dev fits better
The hosted route removes the setup, the GPU and the account.
Result in seconds
Drop a PDF and get clean Markdown with nothing to install – no Python, no GPU, no account.
API + hosted MCP
Call the same conversion from your code or an agent pipeline without managing any infrastructure.
OCR & tables built in
Scanned PDFs, tables and formulas are handled out of the box, no flags to tune.
Occasional or free use
Convert anonymously in the browser on the free tier instead of standing up a GPU box for a few files.
Engine choice
Switch between MinerU and Docling per document instead of tuning one pipeline yourself.
No training on your files
Files are converted and dropped on a short retention window; they are not used to train models.
Converting at scale or from an agent?
pdf2md.dev is a REST API and a hosted MCP endpoint as well as a web app. Convert a PDF and parse clean Markdown from your own code or an LLM agent – no GPU box to run. See the Python tutorial on the developer hub.
Common questions
Is pdf2md.dev a drop-in replacement for Marker?
They are different shapes. Marker is a self-hosted Python library you run yourself; pdf2md.dev is a hosted service you reach from the browser, a REST API or a hosted MCP with nothing to install. Pick by whether you want to host the conversion yourself or have it hosted for you.
Is Marker free?
Marker code is Apache 2.0. Its model weights have a separate modified AI Pubs Open Rail-M licence: the published terms distinguish research, personal use and startups under $5M funding/revenue from other commercial use. Check the current code and weight licences for your use.
Does pdf2md.dev need a GPU?
No. We host the conversion, so you need nothing locally: no GPU, no Python and no install. Marker runs on CPU but is built for a GPU and is much faster on one.
Which is more accurate?
Marker supports an optional --use_llm mode and CPU, GPU or MPS execution. Model configuration, hardware and external LLM costs differ. We have not established an accuracy winner on a shared test corpus; compare the same PDFs with pinned versions and settings.
Can I call either from code or an agent?
pdf2md.dev exposes a REST API and a hosted MCP endpoint, so an agent can convert a PDF without any setup. Marker is a Python library you import in your own code, with an optional paid Datalab API for a hosted endpoint.
What about scanned PDFs?
Both OCR scanned and image-only PDFs. Marker uses Surya OCR when you self-host. On pdf2md.dev OCR is built in across many languages with no flags to set – see converting scanned PDFs.
Is pdf2md.dev really free with no sign-up?
Yes. Convert anonymously in the browser on the free tier (3 slots, 10 MB files, a 15-minute time budget, 1-hour retention). Paid tiers raise every limit.