Every one of these four turns a PDF into Markdown that looks right, which is exactly the problem: a dropped table row and a two-column page read in the wrong order both come out as clean prose, and downstream they become a confident wrong answer carrying a citation that resolves to the correct document. So do not pick on the leaderboard — the only published head-to-head among the four lives in one project's own README and is topped by that project's paid hosted product. Pick on whether the output tells you which page and which box each span came from, because that is the only signal you can gate on.
At a glance
Read from each project's own repository on 1 November 2026. Star counts are a popularity measure and nothing else; they appear here because the ordering is counter-intuitive.
| Project | Licence | Stars | Stated purpose |
|---|---|---|---|
| MarkItDown (Microsoft) | MIT | 193.3k | Lightweight converter for LLM and text-analysis pipelines |
| MinerU (OpenDataLab) | MinerU Open Source License — Apache 2.0 plus conditions | 82.9k | Complex documents to LLM-ready Markdown and JSON |
| Docling (LF AI & Data) | MIT (code); models under their own licences | 69.9k | "Get your documents ready for gen AI" |
| Marker (Datalab) | Apache-2.0 (code); weights under modified AI Pubs Open Rail-M | 40.7k | PDF to Markdown and JSON, quickly and accurately |
The only scoreboard, and who keeps it
Marker's README publishes macro-averaged olmocr-bench results that include its competitors. On that chart Marker balanced scores 76.0% overall and 83.5% on born-digital documents, Marker fast 66.6% and 71.6%, Marker fast with OCR disabled 43.6% and 55.8%, the MinerU pipeline 72.7% overall, and Docling 50.3% overall. Top of the table, at 85.8% overall, is Chandra 2 — a hosted model from Datalab, the company that publishes Marker.
Take that seriously in both directions. The benchmark itself is third-party and the measurements are reproducible in principle, which is more than the other three offer, and nothing here suggests the numbers are wrong. But the chart is assembled by one competitor, run by that competitor, and won by that competitor's commercial product; the version of Marker that scores 43.6% is also Marker. Any of those three facts alone would be reason to treat the ordering as a shortlist rather than a decision. Together they mean the honest use of this chart is "Docling is probably weaker on scanned pages than Marker and MinerU, and I should check that on my own documents," which is a genuinely useful thing to learn and a long way from a ranking.
Note what is not on any chart: MarkItDown, MinerU and Docling publish no accuracy numbers of their own. For three of the four tools in this category there is no vendor claim to distrust, because there is no vendor claim.
MarkItDown — the one that tells you it is not trying
MarkItDown is the most-starred document tool in this space by a factor of 4.7 over Marker, and its README is unusually honest about what that popularity is for. It describes itself as a lightweight utility for LLM and text-analysis pipelines, says its output is "meant to be consumed by text analysis tools", and warns that it may not be suitable for high-fidelity conversion intended for human reading.
What you get
Breadth, and a very short path to a string. It ingests PDF, Word, PowerPoint, Excel, HTML, CSV, JSON, XML, EPub, ZIP archives, images with EXIF and OCR, audio with EXIF and transcription, and YouTube URLs. For an agent that has to accept whatever a user attaches, that list is the product, and no other tool here covers the office long tail as cheaply.
Where it stops
Layout understanding is not in the box. Heavier OCR and layout analysis are delegated to Azure Document Intelligence or Azure Content Understanding — cloud services, billable, and a different trust and data-residency conversation than a local library. The Content Understanding path emits <!-- page 1 --> comment markers, so you can recover a page number there; there are no bounding boxes anywhere, and the repository states no accuracy figures.
The right way to read its star count is that most people converting documents do not have a fidelity problem — they have a "get the text out of fourteen file formats" problem, and MarkItDown solves that one better than anything else here. It is the wrong default for a PDF corpus where the tables matter.
MinerU — tiers, and a licence to actually read
MinerU describes itself as transforming complex documents into LLM-ready Markdown and JSON for agentic workflows, and it is the most configurable of the four. Version 4.x exposes four quality tiers — flash, basic, standard, advanced — and splits the work between small models running on ONNX or Torch and a VLM served by llama.cpp, vLLM or LMDeploy. The repository lists nine rendering targets: Markdown, HTML, LaTeX, DOCX, EPUB, PDF, Structured Content and Content List V1 and V2, each CLI and API exposing its own subset.
Provenance
This is where MinerU earns its place. Its result contract addresses content with locators of the form doc:{short_id}/tier:{tier}/page:{page_no}/block:{block_no}, with page and block numbers 1-based, and block images require a valid bounding box. A locator that names the tier as well as the page is a detail worth noticing: it means a stored span records not just where it came from but how hard the pipeline was trying, which is precisely what you need when you re-process a corpus and want to know which answers to re-check.
The licence
"MinerU Open Source License", described in the repository as based on Apache 2.0 with additional conditions. That phrase is doing real work and it is not Apache 2.0. If you are shipping this inside a product, that sentence is the one to put in front of whoever signs off on dependencies, before you benchmark anything.
Docling — the governance answer
Docling is MIT-licensed code, hosted as a project in the LF AI & Data Foundation, and was started by the AI for knowledge team at IBM Research Zurich. For anyone whose procurement process cares who owns a dependency and what happens if the sponsor loses interest, that combination is not a soft factor — it is the only one of the four with neutral foundation governance, and it is why Docling keeps winning evaluations that Marker wins on score.
What it produces
A DoclingDocument, which the project calls a unified, expressive representation, serialisable to Markdown, HTML, WebVTT, DocLang, DocTags and lossless JSON. Its PDF understanding covers page layout and reading order — the second of which is the failure mode nothing else here advertises handling, and the one that silently destroys a two-column paper. Integrations ship for LangChain, LlamaIndex, CrewAI and Haystack, so it is the least work to drop into an existing retrieval pipeline.
The caveat
Two of them. Docling scores 50.3% on the one published comparison, well behind Marker and MinerU, and that chart is kept by a competitor — but it is the only number anyone has, and it points at scanned and image-heavy documents as the place to test hardest. Separately, the repository landing page does not state whether DoclingDocument records per-element bounding boxes; "lossless JSON" and explicit reading-order handling are promising, and if bbox-level provenance is a requirement you should confirm it in the schema rather than infer it from the word lossless.
Marker — best published score, and a threshold in the licence
Marker converts PDFs to Markdown, JSON, HTML and pre-made chunks, and every format also returns a metadata dictionary with a table of contents and per-page statistics. Its JSON is the most provenance-rich output of the four: each block carries an id that encodes the page (for example /page/10/Page/366), a block_type, a polygon giving four corner coordinates, and section_hierarchy on child blocks. That is enough to render a highlight over the original page, which is the difference between a citation a user can check and a citation a user has to trust.
It also has a hybrid mode: --use_llm sends hard regions to a vision model, with Gemini, Google Vertex, Ollama, Claude, OpenAI-compatible endpoints, Azure OpenAI and OpenRouter supported and gemini-3.5-flash as the stated default. The 76.0% figure is the balanced configuration; the no-OCR fast path scores 43.6%, so the configuration you choose moves the result further than the choice between tools does.
Marker's code is Apache-2.0. Its model weights are not: they ship under a modified AI Pubs Open Rail-M licence, free for research, personal use, and startups under $5M in funding or revenue, with other commercial use requiring a paid licence from Datalab. Crossing $5M in revenue is not a day you want to discover a dependency relicensed itself. This is the single sentence in this post most likely to change a decision.
The axis that actually predicts downstream quality
Here is the failure these tools actually produce in production, and why accuracy percentages do not catch it. A table loses one row to a page break. A two-column page is read straight across, interleaving two unrelated sentences. A footnote marker lands mid-sentence. In all three cases the Markdown is syntactically perfect and reads fluently, so the parse "succeeded"; the chunk embeds fine, retrieval returns it for the right query, and the model answers from it. Your retrieval metrics are green because retrieval was correct. The answer is wrong because the span was.
Nothing in the pipeline downstream of the converter can detect this, because every downstream component's input is the converter's output and it has nothing to compare against. The one thing that makes it detectable is provenance: if each span carries a page and a bounding box, you can render the cited region over the original page and a human spots an interleaved two-column read in under a second. If it carries only a page number, you can at least narrow the check. If it carries nothing, the only way to find the error is for someone to notice the answer was wrong and go looking.
So the ordering that matters is: Marker and MinerU give you element-level locators today, Docling probably can and you should verify, MarkItDown gives you a page comment on one code path. That ordering is almost unrelated to the accuracy chart, and it is the one that determines whether a parsing bug reaches a user as a visible artefact or as a plausible sentence. Document parsing for RAG covers the pipeline this sits in, and citations and source attribution covers what to do with the coordinates once you have them.
When to pick which
| Situation | Pick | Because |
|---|---|---|
| Mixed attachments, text is enough | MarkItDown | Widest format list, MIT, no models to host; it says itself it is not for fidelity |
| Scanned or image-heavy PDFs, tables matter | Marker or MinerU | Best published scores and real per-element locators |
| Procurement cares about governance | Docling | MIT code under LF AI & Data; no single corporate owner |
| You will cite spans back to users | Marker | polygon plus block_type per block renders a highlight |
| Commercial product, revenue over $5M | Not Marker's weights | Paid licence required above the threshold |
| You need to re-process and re-check later | MinerU | Locators record the tier as well as the page and block |
And whichever you pick: the benchmark that decides this is fifty of your own worst pages, scored by you, with the span-level output stored. That takes an afternoon, it is the only measurement in this post that is about your corpus, and every number quoted above is a reason to run it rather than a substitute for it.
FAQ
Which document-to-Markdown converter is most accurate?
On the only published head-to-head — macro-averaged olmocr-bench results in Marker's own README — Marker balanced leads the open tools at 76.0% overall, with the MinerU pipeline at 72.7% and Docling at 50.3%. That chart is published by Marker's vendor and topped by that vendor's paid hosted product, so treat it as a shortlist and verify on your own documents.
Why is MarkItDown so much more popular if it scores nothing?
Because most people are solving a different problem. MarkItDown covers fourteen-plus input formats with an MIT licence and no models to host, which is exactly right when you need text out of whatever a user attached. Its README states it may not suit high-fidelity conversion, and it publishes no accuracy numbers.
Which ones give me bounding boxes?
Marker's JSON gives each block an id encoding the page, a block_type and a polygon of four corners. MinerU addresses blocks with doc/tier/page/block locators and requires a valid bounding box for block images. MarkItDown gives page comment markers on its Content Understanding path only. Docling's landing page does not state per-element bounding boxes — check the DoclingDocument schema if you need them.
Are all four actually free for commercial use?
No. MarkItDown and Docling's code are MIT. Marker's code is Apache-2.0 but its model weights ship under a modified AI Pubs Open Rail-M licence that is free only for research, personal use and startups under $5M in funding or revenue. MinerU uses its own licence, described as Apache 2.0 with additional conditions. Read the last two before you ship.
Do I need a GPU?
For MinerU's higher tiers and Marker's model-based paths, a GPU changes the economics — MinerU's repository claims core inference on supported accelerators runs at roughly two to three times CPU speed. MarkItDown needs none for its built-in converters and pushes heavy OCR to Azure services instead.
Can I just use two of them?
Yes, and it is often the right answer: route office documents and images through MarkItDown for breadth, and PDFs through Marker, MinerU or Docling for structure. Keep one output schema across both so that downstream code does not have to know which converter ran.
Further reading
On this wiki:
- Document parsing for RAG — where this step sits and what it hands the chunker.
- Chunking and vector search — the stage that inherits every parsing error.
- Citations and source attribution UX — turning a bounding box into a checkable citation.
- Hallucination and grounding — why a correct retrieval can still produce a wrong answer.
- Evaluating RAG — building the fifty-page corpus test this post keeps recommending.