Docling vs MinerU vs Marker (2026): Code and Model Licenses

Parser accuracy is something you can measure on your own corpus in a week. A license condition is something you discover in procurement, or worse, after the pipeline has indexed the whole archive and legal asks which model produced them. Re-parsing a production corpus is a re-index, and a re-index forced by a license finding runs on legal's schedule rather than yours.

So the order of operations we recommend is inverted from the usual one: read the terms first, eliminate what your organization cannot run, then benchmark what is left.

One statement, once, so it does not have to be hedged in every paragraph: this is not legal advice. The license text and your counsel decide. What follows quotes the text verbatim, dates it as of September 2026, describes what the words say, and names the clause to put in front of a lawyer. Where a question needs a legal reading, we say so and stop.

Start with what each tool declares and what it pulls. Weight licenses below are the Hugging Face license tags on the default repos, fetched September 2026.

Then the conditions attached to each, which is the table your reviewer actually needs.

One more line belongs under the Docling row. Docling 2.130.0 ships VLM presets that load third-party checkpoints, including datalab-to/chandra-ocr-2 (tagged openrail) and opendatalab/MinerU2.5-Pro-2604-1.2B (apache-2.0). The weight license travels with the checkpoint you select, not with the MIT wrapper that loads it. A Docling deployment is only as permissive as its pipeline configuration.

Marker's README summarizes its weight license as "free for research, personal use, and startups under $5M funding/revenue". That summary is shorter than the clause it describes. Here is Attachment A, section 2 of MODEL_LICENSE, headed "Commercial:", verbatim as of September 2026:

(a) for any purpose if You (your employer, or the entity you are affiliated with) generated more than five million US Dollars ($5,000,000) in gross revenue in the prior year, except where Your Use is limited to personal use or research purposes; > (b) for any purpose if You (your employer, or the entity you are affiliated with) has raised more than five million US dollars ($5,000,000) in total equity or debt funding from any source, except where Your Use is limited to personal use or research purposes; or > (c) for any purpose if You (your employer, or the entity you are affiliated with) provides or otherwise makes available any product or service that competes with any product or service offered by or made available by Licensor or any of its affiliates.

Read textually, three things stand out, and none of them is a legal conclusion:

Three more clauses belong on counsel's desk. Clause 7, "Attribution", requires you, "In connection with any Output", to "give appropriate credit and attribution to Licensor". Clause 8, "Share-a-Like", requires you to apply the license "to any and all copies of the Model, Derivatives of the Model ... and to the Output and any derivatives, changes or improvements to or of the Output." Clause 9, "Updates and Runtime Restrictions", states that the "Licensor reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License". What "Output" covers when the output is Markdown extracted from your own documents, and what remote restriction could mean for an air-gapped install, are questions for a lawyer, not for this post.

A requirements file pinned to marker-pdf 1.10.2 is GPL-3.0 code over weights with $2M triggers. A pin on 2.0.0 is Apache-2.0 code over weights with $5M triggers. The weights in your cache, not the version you think you run, decide which text applies. For use beyond those terms, Datalab sells a commercial license, including an on-prem option, listed on Datalab's pricing page; we have not reviewed its terms.

MinerU's LICENSE.md opens: "MinerU is licensed under Apache License 2.0 and is subject to the additional terms below. Except to the extent expressly modified or supplemented by these additional terms, your other rights and obligations are governed by Apache License 2.0." The additional terms, verbatim as of September 2026:

1. MinerU may be used for commercial purposes without a separate commercial license. However, if you and your Affiliates, on a consolidated basis, meet either of the following thresholds, you must obtain a separate commercial license from [MinerU Team] before continuing such use: a. monthly active users (MAU) exceed 100 million; or b. total monthly revenue exceeds USD 20 million. > 2. If you provide online services to third parties based on MinerU, you must clearly and prominently indicate, in the relevant product or service interface or in publicly available documentation, that MinerU is used. > 3. Where a separate commercial license is required under Section 1 but is not obtained before continuing such use, or where the attribution obligation under Section 2 is not complied with, this License and all rights granted under this License will terminate automatically, and no further notice from the Licensor is required. > 4. In these additional terms, "Affiliates" means any legal entity that directly or indirectly controls, is controlled by, or is under common control with you. "Control" means the power to direct the management and operating decisions of an entity, whether through equity ownership, voting rights, contractual arrangements, or otherwise.

The arithmetic of section 1 is simple and easy to misread. USD 20 million a month is USD 240 million a year, and it is total revenue of you and your Affiliates on a consolidated basis, not revenue attributable to MinerU. A large banking or insurance group clears that figure. Many community banks, credit unions and regional insurers do not, and a subsidiary's own revenue is not the number the clause names. Which side of the line your group sits on is a question for counsel with your consolidated accounts in front of them.

Section 2 is scoped to "online services to third parties". Whether an internal RAG system serving your own employees is such a service is exactly the kind of reading we will not make here. Section 3 is the clause to flag in bold for a reviewer: termination is automatic and requires no notice.

Commits are not what your lockfile installs, so check the PyPI metadata instead. Every release up to and including 3.0.9 (2026-04-07) declares AGPL-3.0 in its License field, although 3.0.0 to 3.0.9 also carry an "OSI Approved :: Apache Software License" classifier, so a scanner that reads classifiers sees a conflict. Every release from 3.1.0 (2026-04-17) onward, including the whole 4.0 line, declares LicenseRef-MinerU-Open-Source-License. No PyPI release was cut during either plain-Apache window, so there is no MinerU version whose License field says plain Apache-2.0.

One dependency note for a PDF-stack license review: MinerU 4.x depends on pypdfium2 (BSD-3-Clause and Apache-2.0), not PyMuPDF, which is dual-licensed under AGPL-3.0 or a commercial license.

Package pins are not weight pins. Every parser here resolves model repos at runtime, and the repos carry their own license tags, which do not always agree with the README you read. For MinerU and Marker, the repos that matter, fetched September 2026:

Three consequences follow.

MinerU's default CPU install pulls the two untagged repos. The 4.0 README says the default install runs small models on ONNX CPU inference and the VLM on llama.cpp in Vulkan mode. That combination fetches MinerU-4_models_onnx, whose card lists PaddlePaddle-derived components (PP-DocLayoutV2, PP-OCRv6, PP-FormulaNet plus-M) and says their original licenses remain applicable, plus the GGUF files from a MinerU2.5-Pro-2605-1.2B-GGUF repo on an individual account outside the opendatalab org, containing the two Q8_0 files and a .gitattributes, with no license tag and no README. The GPU path pulls repos tagged apache-2.0. Same package, same version, different license posture depending on the backend flags.

Marker has no weight-free mode. Even with --disable_ocr in fast mode, the 20M-parameter layout model comes from datalab-to/surya_layout2, which carries the same modified OpenRAIL-M text as the OCR weights, including the $5,000,000 triggers and clause 2(c).

Tags hide modifications. All three Datalab weight repos are tagged just "openrail". A scanner that reads Hugging Face metadata records a standard OpenRAIL. The dollar thresholds and the competitor clause live in the LICENSE file inside each repo, which we verified is the same text as Marker's MODEL_LICENSE apart from whitespace.

The fix is procedural: inventory weights by repo id and revision from the cache you actually staged, not by package name from the requirements file.

Inside an enclave the parser cannot download anything on first run, so staging is where the license inventory gets built as a side effect. The generic Hugging Face offline mechanics and an egress-deny smoke test are covered in our vLLM air-gapped deployment guide; what follows is only what differs per parser.

docling-tools models download -o /opt/models/docling

export DOCLING_ARTIFACTS_PATH=/opt/models/docling
# or per invocation:
docling --artifacts-path=/opt/models/docling report.pdf

mineru-kit models download --tier standard --small-backend onnx --vlm-engine llama-cpp --source huggingface
mineru-kit models verify --tier standard --small-backend onnx --vlm-engine llama-cpp

export MINERU_MODEL_SMALL_BACKEND=onnx
export MINERU_MODEL_VLM_ENGINE=llama-cpp
export MINERU_MODEL_SOURCE=local
mineru-kit parse document.pdf -o document.md --tier standard

export SURYA_INFERENCE_URL=http://127.0.0.1:8000/v1
# or, for the auto-spawned llama.cpp backend:
export SURYA_GGUF_LOCAL_MODEL_PATH=/opt/models/surya/surya-2.gguf
export SURYA_GGUF_LOCAL_MMPROJ_PATH=/opt/models/surya/surya-2-mmproj.gguf
export MODEL_CACHE_DIR=/opt/models/datalab

Format coverage stopped being a differentiator this year. Docling's standard install covers PDF, Office formats, web pages, LaTeX and email. MinerU 4.0 lists PDF, images, DOC/DOCX, PPT/PPTX, XLS/XLSX, RTF, OpenDocument, EPUB, OFD, HTML/MHTML and CSV/TSV, with everything except PDF and images handled by its local Flash native parsing. Marker lists PDF, image, PPTX, DOCX, XLSX, HTML and EPUB with marker-pdf[full].

Hardware defaults differ more. MinerU's base install is CPU-first, with roughly 0.8 GB of models for the basic tier and roughly 2 GB for standard on ONNX and llama.cpp (about 3 GB with vLLM), per its README table, and it recommends mineru[full] on NVIDIA. Marker defaults to balanced mode on GPU and fast mode on CPU, and --disable_ocr needs no inference server at all. Docling runs its default models locally through torch.

On accuracy, the only current numbers in hand are Datalab's own, from the Marker v2.0.0 release notes in July 2026: olmOCR-bench 76.0 in balanced mode, 66.6 in fast mode and 43.6 with OCR disabled, and a claim of more than 5x the pages per second of MinerU's pipeline backend. Those are vendor self-reported and we have not reproduced them. We have no neutral head-to-head of all three at current versions, so we do not rank them.

What decides accuracy is your corpus. Before any parser change reaches a production index, run the shadow-diff loop described in our long-document parsing post (its section "The shadow-diff loop before a production index moves"). Score the parser's structured output where it gets consumed, by the chunker, as covered in chunking for context preservation. If tables and charts carry the answers, consider whether page-image retrieval beats OCR for that slice of the corpus.

This table routes situations to the clause that needs reading. It does not tell anyone they are compliant. Seven situations:

Where our own recommendation lands, for a regulated organization self-hosting a RAG ingestion path: Docling is the one whose default code and weights carry no size trigger, no competitor clause and no termination-without-notice term, and whose offline staging is a documented single command. Use it as the baseline. Evaluate MinerU on a GPU path, where its weights are tagged apache-2.0, once counsel has read sections 1 to 3 against your consolidated figures. Evaluate Marker only with a commercial license in hand or a clear counsel reading of 2(a) through 2(c), because every weight it loads carries those clauses. If your system falls under the EU AI Act high-risk regime, the same repo-plus-revision record belongs in the Annex IV technical file for an open-weight model. For how parsing fits the wider retrieval stack, the rest of our RAG systems guides cover what comes after it.

The action for this week is small. On a connected staging machine, run each candidate parser's model staging for the exact backend you will deploy. List every weight repo and revision that landed in the cache, next to the dated license text from that repo's LICENSE file rather than its tag. Send that list, with the clauses quoted above, to counsel before the first production ingest.