Model licences#

paperextract is Apache-2.0 and distributes no model weights. The backends and their models are separate works with their own licences, which apply to anyone who downloads and runs them: paperextract models fetch downloads from the publishers’ repositories, and using a backend means accepting its model’s terms. This page summarizes what those terms say as of 24 September 2026. It is a summary to help choose a backend, not legal advice; the licence texts in the model repositories govern.

Summary#

Set

Model

Licence

Notes

mineru

opendatalab/MinerU-4_models_onnx

Apache-2.0 (inferred)

The repository declares no licence; its README says the original licences apply, and its sources (PaddlePaddle PP-DocLayoutV2, PP-OCRv6, PP-FormulaNet and MinerU’s torch models) declare Apache-2.0.

mineru

jinzhenj/MinerU2.5-Pro-2605-1.2B-GGUF

Apache-2.0 (inferred)

A third-party quantization with no declared licence, of opendatalab/MinerU2.5-Pro-2605-1.2B, which declares Apache-2.0.

mineru-cuda

opendatalab/MinerU-4_models_torch

Apache-2.0

Declared on the model card.

mineru-cuda

opendatalab/MinerU2.5-Pro-2605-1.2B

Apache-2.0

Declared on the model card.

docling

docling-project/docling-layout-heron

Apache-2.0

Declared on the model card.

docling

docling-project/docling-models

CDLA-Permissive-2.0, Apache-2.0

Declared on the model card.

docling

docling-project/CodeFormulaV2

CDLA-Permissive-2.0

Declared on the model card.

marker

datalab-to/surya-ocr-2-gguf, surya_layout2, OCR error model

AI Pubs OpenRAIL-M, modified

Use restrictions, attribution and share-alike that extend to outputs; see below.

marker

llama.cpp b10964 server

MIT

describe

Qwen/Qwen3.8-27B

Apache-2.0

Licence file shipped with the weights.

The backend code#

  • MinerU (the default backend) is released under the MinerU Open Source License: Apache-2.0 with additional terms. Commercial use needs no separate licence unless you and your affiliates exceed 100 million monthly active users or USD 20 million total monthly revenue, and an online service built on it must give attribution.

  • Docling is MIT.

  • Marker code is Apache-2.0; its Surya models are licensed separately.

  • vLLM, used for GPU inference, is Apache-2.0.

Marker’s Surya models#

The Surya models carry a modified AI Pubs OpenRAIL-M licence (the LICENSE file in each repository). In summary:

  • Use restrictions apply to the model, its derivatives and its output: no use that breaks the law or infringes others’ rights, and no commercial use by an organization (with its employer or affiliates) that had more than USD 5 million in gross revenue in the prior year or raised more than USD 5 million in funding, except personal or research use, and no use by anyone offering a product or service that competes with the licensor’s. Broader licences are offered by the licensor.

  • Attribution: using the output requires crediting the licensor, linking the model and providing the licence.

  • Share-alike: the licence must be applied to copies and derivatives of the model and to the output and its derivatives.

A library extracted with --backend marker therefore carries these terms. MinerU, the default, and Docling do not impose conditions on their output.

Your papers#

Extracted papers remain their publishers’ or authors’ copyrighted works. paperextract keeps them locally; sharing a library is subject to the papers’ own terms.