您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：10月27日，百度发布并开源文心 Unlimited OCR 2，全球首次实现任意页数文档解析，将 OCR 的处理上限从数十页推向“无限”，并达到最先进水平（S
10月27日，百度发布并开源文心 Unlimited OCR 2，全球首次实现任意页数文档解析，将 OCR 的处理上限从数十页推向“无限”，并达到最先进水平（SOTA）。实测显示，该模型能够一次性解析《哈利·波特》全系列 7 部作品，累计超过 3600 页，在文档长度大幅增加的情况下，依然保持较高的精度和处理效率，推动文档解析技术迈向OCR 3.0时代。
公开基准测试显示，在面向多页文档解析的 MPDocBench，以及专门评估百页级超长文档解析能力的 MarathonDocBench 上，文心 Unlimited OCR 2 的综合得分分别为 84.6分和 77.1分，达到端到端性能SOTA，大幅领先 Qwen 3.8-Max、Gemini-3.8-Flash、GLM-5.3-Flash、DeepSeek-4.1-Flash等模型。
这一突破得益于文心 Unlimited OCR 2 所提出的稀疏参考滑动窗口注意力机制（Sparse Reference Sliding Window Attention，SR-SWA）。该机制借鉴了人类抄录书籍的过程——通常专注于当前页面，只有在需要核对前文或查找跨页信息时，才会回看相关内容，而非反复翻阅整本书的内容。
SR-SWA 将该思路引入模型的注意力计算中。具体而言，页面门控（Page Gate）先把每个查询的视觉参考范围从整个文档缩小到相关页面，同时建模跨页的位置关系；随后引入得分最高的视觉 Token 桶（Token Buckets），以捕捉长距离视觉依赖关系。报告测试显示，在百页文档场景中，SR-SWA 的单步解码速度相比 Full Attention 和上一代 R-SWA 分别提升7倍和5.6倍。
随着 AI 技术的发展，长篇研报、整本书籍等超长文档的解析需求日益增长，对 OCR 技术提出了全新挑战。传统 OCR 的解析范围多局限于单页，多页解析通常依赖于结果拼接；第一代 Unlimited OCR 通过引入参考滑动窗口注意力机制，将解析范围拓展至连续数十页；而此次发布的 Unlimited OCR 2 则进一步突破了超长文档解析的计算瓶颈，通过SR-SWA减少冗余视觉信息计算，结合无限解码机制（Unlimited-Horizon Decoding）降低超长文档解析的计算与显存开销，推动文档解析技术迈向OCR 3.0时代，实现了真正意义上的任意页数文档解析。
Unlimited OCR 项目已经获得全球开源社区广泛认可。第一代 Unlimited OCR 于6月22日正式发布后，次日即登顶 GitHub Daily Trending 总榜、Python 榜， HuggingFace 全球模型总趋势榜、多模态模型趋势榜，实现四榜第一，GitHub Star 5 天突破 1 万，跻身现象级爆款项目行列。截至目前，Hugging Face 累计下载量突破 1000 万次，GitHub 斩获 2.85 万 Star，获得图灵奖得主、AI 科学家杨立昆（Yann LeCun）的转发推荐。
11月12日，百度世界2026大会将进一步展示文心 Unlimited OCR 2 的技术亮点及应用潜力。 本届大会以“下一幕，智能组织”为主题，将发布百度“芯云模体”全栈AI的最新升级成果，助力企业和开发者加速AI应用落地，推动智能组织进化。
目前，文心 Unlimited OCR 2 的代码和模型权重同步开源至 GitHub、Hugging Face，并同步上线百度智能云平台和百度AI开放平台，开发者和企业用户可下载使用。
文心 Unlimited OCR 2 开源地址：
- GitHub：https://github.com/baidu/Unlimited-OCR
- HuggingFace：https://huggingface.co/baidu/Unlimited-OCR-2
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: On October 27, Baidupublish and open source Ernie Unlimified OCR 2, for the first time in the world, any page-page document was analysed, moving the OCR processing ceiling from dozens of pages to “unlimited” and reaching the most advanced level (S-B)
On 27 October, Baidupublish and open source Ernie Unlimified OCR 2, for the first time in the world, achieved any page file resolution, pushing the OCR processing ceiling from dozens of pages to “unlimited” and to the most advanced level (SOTA). The results show that the model was able to analyse the entire Harry Potter series of seven works in a single-time manner, cumulatively over 3,600 pages, with high accuracy and processing efficiency in the event of a significant increase in the length of the document, driving the file resolution technology towards the OCR 3.0 age.
Open baseline tests show that in MPDocBench, which addresses multipage documents, and in MarathonDocBench, which specializes in assessing 100-page ultra-long document resolution capabilities, Ernie Unlimied OCR 2 scored a combined score of 84.6 and 77.1 points, respectively, to end-to-end SOTA, leading significantly to models Qwen 3.8-Max, Gemini-3.8-Flash, GM-5.3-Flash, DeepSeek-4.1-Flash.
This breakthrough was the result of a thin reference slide window attention mechanism proposed by Ernie Unioned OCR 2. The mechanism draws on the process of human copying of books — usually focusing on the current page — and only looks back at the text when it needs to be checked or cross-page information, rather than re-reading the contents of the book as a whole.
SR-SWA introduces this idea into the model's attention calculation. Specifically, Page Gate first narrows the visual reference range for each query from the entire document to the relevant page, while modeling the location relationship across the page; it then introduces the highest score to the Token Buckets to capture long distance visual dependency. The report test shows that in the 100-page document scene, the one-step decoded speed of SR-SWA was seven times and 5.6 times higher than that of Full Attention and the previous generation R-SWA, respectively.
With the development of AI technology, there is a growing demand for the resolution of very long documents, such as long articles, books and so on, posing new challenges to OCR technology. The traditional OCR resolution is mostly limited to single pages, with multipage resolution usually relying on a result fusion; the first generation Unlimied OCR expanded the resolution to dozens of consecutive pages by introducing reference slide window attention mechanisms; and this time publish Unlimied OCR 2 further breached the computational bottlenecks of hyper-document resolution by reducing redundant visual information calculations through SR-SWA, and by reducing the calculation of ultra-long document resolution in conjunction with VRAM, facilitating document resolution technology to the OCR 3.0 era, achieving a real level of any page resolution.
The Unlimied OCR project has been widely endorsed by the global open-source community. After the first generation of Unlimided OCR, officially publish on June 22, the next day, GitHub Daily Trending, Python, HuggingFace Global Model General Trends, Multi-temporal Model Trends, achieves the fourth highest, GitHub Star 5-day breakthrough of 10,000, joining the body-grade explosive project. So far, Hugging Face has gained 10 million times in cumulative downloads, and GitHub has won 28.5 million Stars, recommended for transmission by AI Scientist Yann LeCun.
On 12 November, the Baidu World 2026 Congress will further demonstrate the technical bright spots and application potential of Ernie Unlimed OCR 2. The theme of the Congress was “The Next Act, Smart Organization”, and the latest upgrade of the entire publishBaidu “Chip Cloud Moot” AI was used to accelerate the landing of AI applications by aiding firms and developers and to drive the evolution of intelligent organizations.
Currently, Ernie Unlimided OCR 2 ' s code and model weights are synchronized with open sources to GitHub, Hugging Face, and Baidu Smart Cloud Platform and Baidu AI Open Platform, which can be downloaded by developers and business users.
Open source address:
- GitHub：https://github.com/baidu/Unlimited-OCR
- HuggingFace：https://huggingface.co/baidu/Unlimited-OCR-2
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.