zhineng.dev
English
Get started
← All models

文心 Unlimited OCR 2

Published Publisher Baidu

Overview

Baidu released and open-sourced 文心 Unlimited OCR 2 on October 27, 2026. It is a multimodal document-parsing model designed for long, multi-page documents. Baidu’s Qianfan documentation describes it as a mixture-of-experts model with approximately 3.3 billion total parameters and 0.5 billion active parameters. (ai.baidu.com)

Capabilities and uses

  • Recognizes document elements including headings, paragraphs, tables, formulas and images, with layout information and reading order. It can join paragraphs and tables that continue across pages. (ai.baidu.com)
  • Uses Sparse Reference Sliding Window Attention (SR-SWA) and unlimited-horizon decoding to reduce redundant visual computation and memory overhead for long-document parsing. (ai.baidu.com)
  • Potential uses include digitizing books and archives, parsing contracts and financial reports, and preparing structured documents for retrieval-augmented generation. (ai.baidu.com)
  • Baidu reports scores of 84.6 on MPDocBench and 77.1 on MarathonDocBench. These are publisher-reported benchmark results. (ai.baidu.com)

Limitations

Baidu describes the underlying approach as supporting unlimited-horizon parsing and reports a test exceeding 3,600 pages. This is not an unlimited hosted API allowance: the mainland Qianfan API documentation specifies a maximum of 500 cumulative pages per request. It accepts images and PDFs, with additional image-size and Base64 file-size limits. Recognition and cross-page reconstruction should be checked for errors, particularly for consequential documents. (ai.baidu.com)

Pricing

Mainland China, Baidu AI Open Platform document-parsing service: Baidu announced a limited-time pay-as-you-go price of CNY 0.02 per page after a monthly allowance of 2,000 free pages. The allowance refreshes monthly, does not roll over, and is subject to the announcement’s user-certification conditions. This is page-based service pricing; separate input-token and output-token rates for this offer were not specified. The promotion’s end date was not verified. Official pricing announcement, checked November 15, 2026: Baidu commercial-service announcement. (ai.baidu.com)

Baidu’s international API documentation was checked, but an official USD price for this model was not verified. Token-based input and output prices for the Qianfan model API were also not verified. The mainland page-based offer should not be assumed to apply to a different API or region. (cloud.baidu.com)

Official sources