opendatalab/mineru
Transforms complex documents like PDFs and Office docs into LLM-ready markdown/JSON for your Agentic workflows.
How mineru is put together
MinerU is a Python document-parsing engine that converts PDFs, images, DOCX, PPTX, and XLSX into Markdown/JSON using a traditional OCR/layout pipeline, VLM-based backends, and a hybrid of both. It ships multiple entrypoints built on FastAPI/uvicorn (a local REST API 'mineru-api', a GPU-worker 'router' that fans tasks out to local workers or remote upstreams, an OpenAI-compatible VLM server) plus a Gradio web UI and Click CLIs. Requests flow through multipart form validation (mineru.cli.api_request) into backend analyzers under mineru/backend/*, which produce 'middle JSON' that is rendered to Markdown/content lists; results are stored as files under an output root and returned inline or as ZIP. Configuration is almost entirely environment-variable driven, with model weights pulled from HuggingFace/ModelScope at runtime.
Languages
Frameworks
Infrastructure
Major components
mineru.cli
User-facing entrypoints: Click CLI client, FastAPI REST service (fast_api.py), request router/load-balancer over local GPU workers and remote upstreams (router.py), Gradio demo app, and OpenAI-compatible VLM servers.
mineru.backend.pipeline
Traditional multi-model OCR/layout/table/formula pipeline that batch-analyzes PDF pages into middle JSON and renders final content.
mineru.backend.vlm + hybrid
Vision-language-model backends (local engine, HTTP client against OpenAI-compatible servers) and the hybrid pipeline+VLM mode with medium/high effort levels.
mineru.backend.office
Native DOCX/PPTX/XLSX parsing into middle JSON and Markdown, including chart/image extraction and mkcontent rendering.
mineru.model
Model implementations and converters: layout detection (PP-DocLayoutV2), OCR (PaddleOCR-style PyTorch port, seal detection), table recognition (SLANet-plus, UNet table), formula recognition (Unimernet, PP-FormulaNet), and office format converters.
mineru.data
Abstractions for reading/writing input/output data across local filesystem, HTTP, and S3 (including multi-bucket S3), plus path/schema utilities.
mineru.utils
Shared utilities: config/env reading, OCR language handling, PDF image/text tools, bbox drawing, hashing, device/system checks, and CLI argument parsing.
docker/
Container images for global and China mirrors targeting many domestic AI accelerator platforms, plus a compose file for deployment.
A new release shipped with improved Unicode handling so PDF text extraction no longer breaks on certain characters like emojis and non-Latin symbols.
2 weeksWeek of 2026-08-10
Over the past two weeks, the MinerU project focused on polishing how text and tables are handled in document conversion. The main highlight was a new feature that fixes garbled characters (like emojis and symbols) when extracting text from PDFs, followed by a smaller fix ensuring special characters inside tables render correctly.
Week by week
2026-08-10A new release shipped with improved Unicode handling so PDF text extraction no longer breaks on certain characters like emojis and non-Latin symbols.latest3 changes
Feature
Better handling of special characters in PDF text extraction
The tool now correctly reassembles split character pairs (used for emojis and some international scripts), so extracted PDF text displays properly instead of showing broken or missing symbols.
Fix
Fixed table rendering for special characters in Word documents
Special characters appearing in table signatures are now rendered correctly when converting documents to Word format.
Chore
New version release
The project's version number was updated and the latest changes were merged into the main codebase for release.
2026-08-03A targeted fix made sure unusual characters inside tables display consistently during document conversion.1 change
Fix
Consistent rendering of special characters in tables
Characters that previously displayed inconsistently in table signatures now render the same way across different output formats.
Dependencies and code review
Dependency advisories
Security Watch
No known advisories across 0 scanned dependencies.
No known advisories in the scanned dependencies.
Code review
No concrete code-level issues confirmed.
Get this report every week for your repos.
GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.
$19 a month, flat · First 10 outputs free · No card required