MinerU
Document-to-Markdown engine from Shanghai AI Lab's OpenDataLab. Relicensed off AGPLv3 in April 2026, reported at 95.4 on OmniDocBench v1.6 and running on ten domestic AI chip families.
Overview
MinerU comes out of OpenDataLab at Shanghai AI Laboratory, where its original job was turning PDFs into clean pretraining corpora at scale. That heritage shows in the output. It takes PDF, DOCX, PPTX, XLSX, images and web pages and produces Markdown or structured JSON with cross-page tables stitched back together, truncated paragraphs rejoined and charts parsed, rather than a flat dump of text boxes in whatever order the PDF happened to store them. Version 3.4 shipped on 18 June 2026 and roughly doubled OCR throughput on the batch path.
There are two routes through the tool and picking correctly is the most consequential decision you will make with it. The pipeline backend uses PP-OCRv6 plus classical layout models, runs on CPU or 4GB of VRAM and is reported at 86.47 on OmniDocBench v1.6. The hybrid and VLM backends run MinerU2.5-Pro-2605-1.2B, want 8GB of VRAM and an NVIDIA GPU from Volta onward, and are reported at 95.39. That is a nine-point spread hiding behind a configuration flag.
Licensing is the 2026 story for this project. MinerU ran on AGPLv3 until 3.1.0, which is why plenty of enterprise architecture reviews rejected it on sight. It now uses the MinerU Open Source License: Apache 2.0 plus two additional terms, namely a separate commercial licence above 100 million MAU or 20 million USD monthly revenue, and a requirement that online services built on it disclose the fact in the interface or public documentation. On deployment, MinerU reaches further past CUDA than anything else in this category, listing Ascend, Cambricon, Kunlunxin, Moore Threads, Hygon, Biren, Enflame, MetaX, Iluvatar and T-Head, with Apple Silicon MPS for local development on a Mac. Weights default to Hugging Face with an environment variable to switch the source to ModelScope, and that is the first switch to flip if you are installing from inside the mainland.
Key Features
- ✓ Converts PDF, DOCX, PPTX, XLSX, images and web pages into Markdown or structured JSON aimed at RAG and agent pipelines
- ✓ Two backends: a pipeline built on PP-OCRv6 that runs on CPU or 4GB VRAM, and a hybrid/VLM path running MinerU2.5-Pro-2605-1.2B that needs 8GB VRAM
- ✓ Reported 95.39 on OmniDocBench v1.6 for the hybrid backend against 86.47 for the pipeline backend
- ✓ 109-language OCR with cross-page table merging, truncated paragraph joining, chart parsing and recognition of images embedded inside tables
- ✓ Runs on Ascend, Cambricon, Kunlunxin, Moore Threads, Hygon, Biren, Enflame, MetaX, Iluvatar and T-Head silicon, plus Apple Silicon via MPS and NVIDIA from Volta onward
- ✓ Ships CLI, REST API, Docker image, WebUI and an MCP server, with LangChain, Dify and FastGPT integrations
Where it holds
- • The April 2026 move off AGPLv3 removes the copyleft problem that had MinerU banned outright at a lot of companies
- • Domestic-chip coverage is the widest of any document stack here, so it is genuinely deployable on Ascend or Kunlunxin hardware rather than CUDA only
- • The pipeline backend runs on CPU, so a low-volume on-premises deployment needs no GPU at all
- • Output is structured for retrieval, not just extracted: cross-page tables and split paragraphs get rejoined, which is exactly the work a naive PDF-to-text step leaves you to do by hand
Where it breaks
- • The additional licence terms are real and specific: over 100 million MAU or 20 million USD monthly revenue triggers a separate commercial licence, and any public service built on it must disclose MinerU by name
- • Weights default to Hugging Face, so a mainland deployment needs the ModelScope environment variable set; it is documented but easy to miss and the first install then hangs
- • The gap between pipeline at 86.47 and hybrid at 95.39 on OmniDocBench v1.6 is nine points, so leaving the backend on a default you did not choose costs real accuracy
- • Release cadence is fast enough to be disruptive: 3.1.0 in April 2026 and 3.4 by June, with model names changing between them
My Take
Legal teams that had this on a blocklist can look again. MinerU moved off AGPLv3 with version 3.1.0 in April 2026 onto a custom Apache 2.0 derivative. Two conditions survive: a separate commercial licence is required above 100 million monthly active users or 20 million USD monthly revenue, and any online service built on it must say so visibly in the product or its docs. Neither threshold touches most teams, but read the clause before you ship a SaaS on it.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- MinerU Open Source License (Apache 2.0 with additional commercial-scale and attribution terms)
- Starting at:
- Free to self-host, including for commercial use, under the MinerU Open Source License. A separate commercial licence must be obtained from the MinerU team once you and your affiliates exceed 100 million monthly active users or 20 million USD in monthly revenue on a consolidated basis. Any online service built on MinerU must visibly disclose that fact in the product interface or public documentation. OpenDataLab also runs a hosted service at mineru.net for teams that would rather not operate a GPU.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where MinerU holds and where it breaks. The fastest check is your own workload.
Visit website ↗MinerU compared head to head
Alternatives to MinerU
dots.ocr
open-sourceRedNote's (小红书) MIT-licensed document parser. A single vision-language model does layout, reading order and text; dots.mocr at 3B added direct SVG output for charts in March 2026.
PaddleOCR
open-sourceBaidu's Apache 2.0 OCR toolkit (飞桨 PaddleOCR). PP-OCRv6 ships at 1.5M, 7.7M and 34.5M parameters, and PaddleOCR-VL-1.6 is a 0.9B document parser covering 109 languages.
RapidOCR
open-sourcePaddleOCR's models converted to ONNX and shipped without the PaddlePaddle framework. Apache 2.0, a 27.3 MB wheel, PP-OCRv6 small by default and fully offline CPU deployment.
LlamaIndex
freemiumOpen-source Python and TypeScript framework for RAG and agentic workflows, paired with LlamaCloud for document parsing, extraction and managed indexes.