mineru-3.0.0-released

## What's Changed

- 2026/03/29 3.0.0 Released

  This release delivers a systematic upgrade centered on **parsing capability, system architecture, and engineering usability**. The main updates include:
  
  - Native `DOCX` parsing
    - Official support for native `DOCX` parsing, delivering high-precision results without hallucinations.
    - Compared with the traditional workflow of first converting `DOCX` to `PDF` and then parsing it, end-to-end speed is improved by tens of times, making it better suited for scenarios with high requirements for both accuracy and throughput.
  - `pipeline` backend upgrade
    - The `pipeline` backend achieves a score of `86.2` on OmniDocBench (v1.5), surpassing the accuracy of the previous-generation mainstream VLM `MinerU2.0-2505-0.9B`.
    - Added support for parsing images/formulas inside tables, seal text recognition, vertical text support, and interline formula numbering recognition, continuously improving parsing quality for complex document scenarios.
    - While maintaining high accuracy, it keeps resource usage extremely low and continues to support inference in pure CPU environments.
  - `API / CLI / Router` orchestration upgrade
    - `mineru` now runs as an orchestration client based on `mineru-api`; when `--api-url` is not provided, it will automatically start a local temporary service.
    - `mineru-api` adds a new asynchronous task endpoint `POST /tasks`, supporting task submission, status querying, and result retrieval; meanwhile, it retains the synchronous parsing endpoint `POST /file_parse` for compatibility with legacy plugins.
    - Added `mineru-router`, designed for unified entry deployment and task routing across multiple services and multiple GPUs; its interfaces are fully compatible with `mineru-api` and support automatic task load balancing.
  - Deployment and usability improvements
    - Resolved compatibility issues with `torch >= 2.8`; the base image has been upgraded to `vllm0.11.2 + torch2.9.0`, unifying installation paths across different Compute Capabilities.
    - Optimized the parsing pipeline with a sliding-window mechanism, significantly reducing peak memory usage in long-document scenarios, so documents with tens of thousands of pages no longer need to be split manually.
    - Batch inference in `pipeline` now supports streaming writes to disk, allowing completed parsing results to be written out in time and further improving the experience for long-running tasks.
    - Completed thread-safety optimization and now fully supports multi-threaded concurrent inference; together with `mineru-router`, this enables one-click multi-GPU deployment and makes it easy to build high-concurrency, high-throughput parsing systems.
    - Completely removed the use of two AGPLv3 models (`doclayoutyolo` and `mfd_yolov8`) and one CC-BY-NC-SA 4.0 model (`layoutreader`).  
  
  This update is not just a set of feature enhancements, but a key leap forward in MinerU's overall system capabilities. We specifically addressed the peak memory usage issue in long-document parsing. Through optimizations such as sliding windows and streaming writes to disk, ultra-long document parsing has moved from “requiring manual splitting and careful handling” to being “stable, scalable, and ready for production workloads.” At the same time, we completed thread-safety optimization and fully enabled multi-threaded concurrent inference, further improving single-machine resource utilization and runtime stability under high-concurrency workloads. On top of this, with `mineru-router` and the new `API / CLI` orchestration framework, MinerU now supports one-click multi-GPU deployment, unified access across multiple services, and automatic task load balancing, significantly reducing the difficulty of large-scale deployment. As a result, MinerU is evolving from a standalone data production tool into a large-scale document parsing foundation for high-concurrency and high-throughput scenarios, providing enterprise-grade document data processing with infrastructure that is more stable, more efficient, and easier to scale.


- 2026/03/29 3.0.0 发布

  本次版本更新围绕**解析能力、系统架构与工程可用性**进行了系统升级。主要更新内容包括：
  
  - `DOCX` 原生解析
    - 正式支持 `DOCX` 原生解析，在无幻觉前提下实现高精度解析。
    - 相较于“先将 `DOCX` 转为 `PDF` 再解析”的传统流程，端到端速度提升数十倍以上，更适合对精度与吞吐均有要求的场景。
  - `pipeline` 后端升级
    - `pipeline` 后端在 OmniDocBench (v1.5) 上取得 `86.2` 分，精度超过上一代主流 VLM `MinerU2.0-2505-0.9B`。
    - 新增表格内图片/公式解析、印章文字识别、竖排文本支持、行间公式序号识别等能力，持续提升复杂文档场景下的解析效果。
    - 在保持高精度的同时，资源占用极低，并继续支持纯 CPU 环境推理。
  - `API / CLI / Router` 编排升级
    - `mineru` 现作为基于 `mineru-api` 的编排客户端运行；在未传入 `--api-url` 时，会自动拉起本地临时服务。
    - `mineru-api` 新增异步任务接口 `POST /tasks`，支持任务提交、状态查询与结果获取；同时保留同步解析接口 `POST /file_parse`，以兼容老版本插件。
    - 新增 `mineru-router`，适用于多服务、多 GPU 的统一入口部署与任务路由；其接口与 `mineru-api` 完全兼容，并支持任务自动负载均衡。
  - 部署与使用体验优化
    - 解决了 `torch >= 2.8` 的兼容问题，基础镜像升级为 `vllm0.11.2 + torch2.9.0`，统一了不同 Compute Capability 的安装路径。
    - 通过滑动窗口优化解析链路，显著降低长文档场景下的内存峰值占用，上万页文档解析不再需要手动拆分。
    - `pipeline` 的 batch 推理支持流式落盘，已完成的解析结果可及时写出，进一步提升长任务处理体验。
    - 完成线程安全优化，全面支持多线程并发推理；配合 `mineru-router`，可一键实现多卡部署，轻松构建高并发、高吞吐解析系统。
    - 完全移除了两个 AGPLv3 模型（`doclayoutyolo` 和 `mfd_yolov8`）以及一个 CC-BY-NC-SA 4.0 模型（`layoutreader`）的使用。  
  
  本次更新不仅是若干功能点的补强，更是 MinerU 在系统能力上的一次关键跃迁。我们重点解决了长文档解析过程中的内存峰值占用问题，通过滑动窗口、流式落盘等链路优化，让超长文档解析从“需要手动拆分、谨慎处理”走向“稳定可跑、规模可扩展”。同时，我们完成了线程安全优化，全面支持多线程并发推理，进一步提升了单机资源利用率与高并发场景下的运行稳定性。在此基础上，基于 mineru-router 与全新的 API / CLI 编排体系，MinerU 已具备一键多卡部署、多服务统一接入、任务自动负载均衡的能力，显著降低了大规模部署难度。至此，MinerU 正在从单一的数据生产工具，进一步演进为面向高并发、高吞吐场景的大规模文档解析基座，为企业级文档数据处理提供更稳定、更高效、更易扩展的基础设施能力。


## New Contributors
* @boshi91 made their first contribution in https://github.com/opendatalab/MinerU/pull/4523
* @Niujunbo2002 made their first contribution in https://github.com/opendatalab/MinerU/pull/4662

**Full Changelog**: https://github.com/opendatalab/MinerU/compare/mineru-2.7.6-released...mineru-3.0.0-released