github opendatalab/MinerU mineru-4.0.0-released
on GitHub

9 hours ago

What's Changed

  • 2026/09/16 4.0.0 Released

    MinerU 4.0 is a major release that rebuilds the project around tiered parsing quality, native multi-format document parsing, a local document library for agent reading, and unified SDK / API / service tools. The main updates include:

    • Four parsing tiers

      • The 3.x pipeline / vlm / hybrid backend selection is replaced by four quality tiers: flash, basic, standard, and advanced.
      • flash parses native PDF text and digital documents without inference models, for discovery, preview, and indexing. basic runs small models for OCR, formulas, and tables, and can run on CPU. standard (the default reading quality) combines small models with the VLM for complex layouts. advanced spends more inference compute for maximum quality on difficult documents.
      • PDF and images support all four tiers; Office, OpenDocument, EPUB, OFD, HTML, and CSV/TSV are parsed locally at flash tier.

    • Native multi-format parsing

      • Supported inputs now cover PDF, images, DOC/DOCX, PPT/PPTX, XLS/XLSX, RTF, ODT/ODS/ODP, EPUB, OFD, HTML, and CSV/TSV. DocVortex provides native document parsing: no conversion to PDF, no model downloads.
      • Plain-text files (.txt, .md, ...) are read directly instead of parsed.

    • Local document library and agent reading

      • New mineru command family: parse, read, find, search, show, list, watch, scan, plus cache control (invalidate, forget, cleanup) and a background service (mineru server start / status).
      • mineru parse defaults to the first 10 PDF pages and returns a continuation command (next_request), so long documents are read progressively page by page instead of loaded at once.
      • Stable locators such as doc:{id}/tier:{tier}/page:{page}/block:{block} support follow-up reads, page/block image export, and verifiable citations. The README embeds an Agent Guide covering these workflows.
      • Privacy first: parsing is local by default, and documents are uploaded only when --remote is passed explicitly; anonymous telemetry can be inspected and disabled (mineru telemetry status).

    • Independent model configuration

      • Small models (ONNX or Torch) and the VLM engine (llama.cpp, vLLM, or LMDeploy) are now configured independently via model.small_backend and model.vlm.engine; model.stack / --stack are removed.
      • The base package works out of the box: ONNX CPU small models plus a llama.cpp VLM in Vulkan mode. mineru[torch] adds Torch small-model dependencies; mineru[full] also installs vLLM (Linux) or LMDeploy (Windows); Apple Silicon installs Torch by default and runs the VLM with llama.cpp.
      • New 4.0 model bundles MinerU-4_models_onnx / MinerU-4_models_torch (PP-DocLayoutV2, PP-OCRv6 Tiny Det + Small Rec, PP-FormulaNet plus-M, seal OCR, and shared table models), downloaded and verified with mineru-kit models download / verify.

    • Unified tools and V1 API

      • The Python SDK, stateless batch conversion (mineru-kit parse), the multi-service Router, and the Gradio WebUI (mineru-webui) share one structured result model.
      • The HTTP service is rebuilt around V1 (/v1/health, /v1/tiers, /v1/uploads, /v1/parse/jobs, /v1/files); legacy /file_parse and /tasks are not provided.

    • Structured results and nine rendering targets

      • One document model renders to Markdown, HTML, LaTeX, DOCX, EPUB, PDF, Structured Content, and Content List V1/V2; each entrypoint exposes its own subset of exports. See Output Formats and Result Contract.
      • Results carry a versioned shared schema, and PDF page ranges (1-5,8,r3-r1; r1 is the last page, all selects the whole document) behave consistently across CLI, library, API, and SDK.

    • Breaking changes and migration

      • Entrypoints: mineru -p input.pdf -o output becomes mineru parse input.pdf --pages all -o output.md (library) or mineru-kit parse input.pdf -o output.md (one-off conversion); mineru-gradio becomes mineru-webui; -m all becomes mineru-kit models download --tier standard.
      • Extras and runtime: mineru[core] / mineru[pipeline] are replaced by the base package, mineru[torch], and mineru[full]; Python >=3.10,<3.15.
      • Configuration now lives in $MINERU_HOME/config.yaml; the old mineru.json is no longer read.
      • Legacy result JSON is auto-converted only for cached Doclib history; old artifacts saved elsewhere should be re-parsed.
      • Existing AMD and vendor accelerator adaptations remain on mineru<4.
      • See the 3.x → 4.0 migration guide for the full list.

    With 4.0, MinerU grows from a document conversion tool into a privacy-first document platform: tiered quality, native multi-format parsing, a searchable local library with stable citation locators for agents, and unified interfaces. Get started with the Quick Start, Tiers and runtimes, and SDK and API pages; the full release history is in the changelog.

  • 2026/09/16 4.0.0 发布

    MinerU 4.0 是一次大版本重构，围绕 分档解析质量、原生多格式文档解析、面向 Agent 阅读的本地文档库 与 统一的 SDK / API / 服务工具 重新打造。主要更新内容包括：

    • 四档解析质量（tier）

      • 3.x 的 pipeline / vlm / hybrid 后端选择升级为 flash、basic、standard、advanced 四档质量档位。
      • flash 原生解析 PDF 文本层与电子文档，无需推理模型，适合发现、预览与索引；basic 使用小模型完成 OCR、公式与表格识别，可在 CPU 环境运行；standard（默认阅读质量）结合小模型与 VLM 处理复杂版面；advanced 在标准能力上投入更多推理算力，应对高难度文档。
      • PDF 与图片支持全部四档；Office、OpenDocument、EPUB、OFD、HTML、CSV/TSV 在本地以 flash 档解析。

    • 原生多格式解析

      • 支持的输入覆盖 PDF、图片、DOC/DOCX、PPT/PPTX、XLS/XLSX、RTF、ODT/ODS/ODP、EPUB、OFD、HTML 与 CSV/TSV；DocVortex 提供原生文档解析，无需转 PDF、无需下载模型。
      • 纯文本文件（.txt、.md 等）直接读取，不做解析。

    • 本地文档库与 Agent 阅读

      • 新增 mineru 命令族：parse、read、find、search、show、list、watch、scan，以及缓存管理（invalidate、forget、cleanup）与后台服务（mineru server start / status）。
      • mineru parse 默认读取 PDF 前 10 页并返回续读命令（next_request），长文档按页渐进阅读，无需一次性载入。
      • doc:{id}/tier:{tier}/page:{page}/block:{block} 等稳定定位符支持后续读取、页/块图片导出与可验证引用；README 内置 Agent Guide 描述这些工作流。
      • 隐私优先：默认全程本地解析，只有显式使用 --remote 才会上传文档；匿名遥测可随时查看与关闭（mineru telemetry status）。

    • 模型配置独立化

      • 小模型（ONNX / Torch）与 VLM 引擎（llama.cpp / vLLM / LMDeploy）通过 model.small_backend 与 model.vlm.engine 独立配置；移除 model.stack 与 --stack。
      • 基础包开箱即用：小模型走 ONNX CPU，VLM 使用 llama.cpp Vulkan 模式；mineru[torch] 增加 Torch 小模型依赖，mineru[full] 额外安装 vLLM（Linux）或 LMDeploy（Windows）；Apple Silicon 默认安装 Torch 并以 llama.cpp 运行 VLM。
      • 新的 4.0 模型包 MinerU-4_models_onnx / MinerU-4_models_torch（PP-DocLayoutV2、PP-OCRv6 Tiny Det + Small Rec、PP-FormulaNet plus-M、印章识别与共享表格模型），使用 mineru-kit models download / verify 下载与校验。

    • 统一工具链与 V1 API

      • Python SDK、无状态批量转换（mineru-kit parse）、多服务 Router 与 Gradio WebUI（mineru-webui）共享同一套结构化结果。
      • HTTP 服务重构为 V1（/v1/health、/v1/tiers、/v1/uploads、/v1/parse/jobs、/v1/files）；不再提供旧的 /file_parse 与 /tasks。

    • 结构化结果与九种渲染目标

      • 一套文档模型渲染为 Markdown、HTML、LaTeX、DOCX、EPUB、PDF、Structured Content 与 Content List V1/V2 九种输出，各入口暴露各自的导出子集，详见输出格式与结果契约。
      • 结果携带带版本号的共享 schema；PDF 页范围（1-5,8,r3-r1，r1 为最后一页，all 表示整本）在 CLI、文档库、API 与 SDK 间行为一致。

    • 破坏性变更与迁移

      • 入口变化：mineru -p input.pdf -o output 变为 mineru parse input.pdf --pages all -o output.md（文档库）或 mineru-kit parse input.pdf -o output.md（一次性转换）；mineru-gradio 变为 mineru-webui；-m all 变为 mineru-kit models download --tier standard。
      • 依赖分组：mineru[core]、mineru[pipeline] 等 extras 调整为基础包、mineru[torch]、mineru[full]；Python 要求 >=3.10,<3.15。
      • 配置迁移至 $MINERU_HOME/config.yaml，不再读取旧的 mineru.json。
      • 旧版结果 JSON 仅在 Doclib 缓存读取边界内自动转换，其余场景请重新解析源文档。
      • 既有的 AMD 及厂商加速卡适配仍停留在 mineru<4。
      • 完整变化清单请参考 3.x → 4.0 迁移指南。

    在 4.0 版本，MinerU 从文档转换工具升级为隐私优先的文档平台：分档解析质量、原生多格式解析、面向 Agent 的可检索本地文档库与稳定引用定位，以及统一的接口体系。欢迎从快速开始、档位与运行时与 SDK 与 API 入手，完整版本历史见 changelog。

Full Changelog: mineru-3.4.5-released...mineru-4.0.0-released

Check out latest releases or
releases around opendatalab/MinerU mineru-4.0.0-released

Don't miss a new MinerU release

NewReleases is sending notifications on new releases.

Get notifications