What is MinerU?
MinerU converts PDFs, scans and images, Word, PowerPoint, Excel, EPUB, OFD, HTML and CSV files into Markdown, HTML, LaTeX, DOCX or structured JSON, keeping tables, formulas and reading order. It is built by OpenDataLab, which also maintains the OmniDocBench benchmark, and its vision model, MinerU2.5-Pro, ships through it.
Version 4.0, released on 16 September 2026, reorganised it around four tiers: Flash reads a PDF's own text layer, Basic adds small OCR, formula and table models that run on CPU, and Standard and Advanced add the 1.2-billion-parameter vision model served by llama.cpp, vLLM or LMDeploy. It also added a local document library that agents can search and cite by page and block. Version 4.1 brought automatic GPU detection across Intel, AMD, NVIDIA and Apple hardware and a new asynchronous Tasks API.
The licence changed in April 2026. Version 3.1.0 replaced AGPLv3 with the MinerU Open Source License, Apache 2.0 plus terms that require a commercial licence above 100 million monthly active users or $20 million in monthly revenue, and require services built on it to state that they use MinerU.
Is MinerU worth it?
MinerU's model is near the top of the field, and the evidence is better than self-reporting alone. OpenDataLab reports 95.69 for MinerU2.5-Pro on OmniDocBench v1.6, up 2.71 points on the same 1.2B architecture through data work alone. The caveat is that OpenDataLab maintains OmniDocBench and defined v1.6 in the same paper. A rival's table, in TII's Falcon OCR model card, is more modest but still strong: 93.22 on OmniDocBench, behind PaddleOCR-VL-1.6 at 94.95, and 81.5 on olmOCR-bench, behind only dots.mocr, Chandra OCR 2 and Falcon OCR itself in that table.
The licence needs reading carefully. The April 2026 move away from AGPLv3 removed the copyleft obligation, and the thresholds, 100 million monthly users or $20 million monthly revenue, will not touch most companies. But any online service built on MinerU must say so prominently, and failing to do that ends the licence automatically. It is not an OSI-approved licence, and some legal teams will treat it like one they have to review. The history adds to the caution: the repository switched from AGPL to Apache and back again in March 2026 before settling in April.
The tiers are a real trade-off. Flash and Basic run on ordinary CPUs, and on a CPU-only machine the Standard tier is enabled automatically only with 12 or more logical cores; the vision-model tiers want a GPU, with 8 GB of VRAM described as a planning starting point rather than a guarantee. The model card reports 2.12 pages a second on one A100 with vLLM. Marker's maintainers measured MinerU's older pipeline at 72.7 on olmOCR-bench and 0.54 pages a second on a B200, so results depend heavily on which tier and version is run.
The project moves fast. There were 18 releases between 4.0.0 on 16 September and 4.1.4 on 2 November, and 4.1 broke the HTTP API, requiring servers, routers, SDKs and clients to be upgraded together. With 82,996 stars and only 11 open issues, maintenance is clearly active; the cost is that pinning a version and reading the migration guide are part of running it.
Where it earns its place
An ML engineer building retrieval over 500,000 Chinese and English research papers
Turn PDFs with formulas and tables into clean Markdown
Run mineru-kit parse at the Standard tier with vLLM on one GPU host. At the model card's 2.12 pages a second on an A100, 5 million pages (ten per paper) take about 27 days on one card, so shard across GPUs with the multi-service Router.
A platform team giving coding agents access to internal documents
Let agents read long documents page by page and cite exact locations
Install the mineru skill and start the local document library; agents parse once, then read by doc, page and block locators instead of loading whole files. Documents are not uploaded unless remote parsing is configured.
A legal-operations lead at a SaaS company embedding a parser in its product
Check the licence before shipping MinerU to customers
Below the 100 million user and $20 million monthly revenue thresholds no commercial licence is needed, but the product's interface or public documentation must state that it uses MinerU, or the licence terminates.
Before you commit
- Setup
- Python 3.11 to 3.14: uv pip install -U "mineru>=4.0,<5", then mineru-kit parse document.pdf -o document.md --tier standard. Use mineru[full] for high-throughput NVIDIA serving; models download from Hugging Face or ModelScope.
- Integrations
- CLI, Python SDK (DoclibClient), V1 HTTP and Tasks APIs, a multi-service Router, a Gradio web UI, an agent skill, and a remote VLM endpoint option; nine output formats including Markdown, HTML, LaTeX, DOCX and JSON content lists.
- Data and privacy
- Runs locally; the docs say documents are not uploaded to the official service unless remote parsing is explicitly configured. MinerU2.5-Pro weights are Apache 2.0; the toolkit is under the MinerU Open Source License.
- Cost at scale
- Self-hosted cost is hardware: Flash and Basic run on CPUs, the docs advise at least 16 GB of system memory for higher-throughput use, and MinerU2.5-Pro reached 2.12 pages a second on one A100 with vLLM.
- Limits
- Automatic tier selection does not check GPU memory; Docker for non-NVIDIA devices is pending an update; Office and web formats use Flash only; HTTP APIs changed incompatibly in 4.1.
What’s good
- MinerU2.5-Pro scores 93.22 on OmniDocBench in a rival's (TII's) comparison
- Four tiers, from CPU text-layer parsing to GPU vision-model parsing
- Office, EPUB, OFD, HTML and CSV parsed natively
- Very active: 18 releases from mid-September to early November 2026
What’s not
- Custom licence with an attribution duty and revenue thresholds
- OpenDataLab also runs OmniDocBench, its headline benchmark
- Breaking HTTP API change in 4.1
- Licence file flipped between AGPL and Apache three times in March and April 2026
Best for
RAG and data teams parsing large, mixed collections of PDFs and Office files on their own hardware, who can accept an attribution clause.
MinerU alternatives
MinerU: common questions
- Is MinerU worth it?
- MinerU pairs one of the most accurate open document models with a busy, well-documented toolkit, though its licence is no longer plain open source and its APIs keep moving. We rate it 4.3 out of 5.
- How much does MinerU cost?
- Free under the MinerU Open Source License (Apache 2.0 with added terms); a separate commercial licence is required above 100 million monthly active users or $20 million monthly revenue.
- What are the downsides of MinerU?
- Custom licence with an attribution duty and revenue thresholds; openDataLab also runs OmniDocBench, its headline benchmark; breaking HTTP API change in 4.1; licence file flipped between AGPL and Apache three times in March and April 2026.
- Who is MinerU best for?
- RAG and data teams parsing large, mixed collections of PDFs and Office files on their own hardware, who can accept an attribution clause.
- How much does MinerU cost for a team or in production?
- Self-hosted cost is hardware: Flash and Basic run on CPUs, the docs advise at least 16 GB of system memory for higher-throughput use, and MinerU2.5-Pro reached 2.12 pages a second on one A100 with vLLM.
- Is MinerU safe for company data?
- Runs locally; the docs say documents are not uploaded to the official service unless remote parsing is explicitly configured. MinerU2.5-Pro weights are Apache 2.0; the toolkit is under the MinerU Open Source License.
- What are the best MinerU alternatives?
- The closest alternatives we have reviewed are Docling, Marker and Unstructured.
Answers checked Nov 2, 2026.
Make MinerU? Show this rating on your site
