Hugging Face researcher Daniel van Strien benchmarked eight open OCR models on 850 pages of historical books and archives, finding the cost to transcribe 1,000 pages fell from about $0.51 with April 2025's RolmOCR to under $0.1 with October 2026's LightOnOCR-3-0.8B, putting a million pages at roughly $65-130 in GPU rental. The study also notes that pipeline optimizations such as balanced page scheduling, smaller scan images and software upgrades can match or exceed the savings from a newer model generation.
OCR
光学字符识别
Technology for recognizing and extracting text from scanned images, central to large-scale archival digitization.
Curated Intelligence
Hugging Face's Daniel van Strien benchmarked eight open OCR models on identical GPU hardware, finding per-1,000-page cost fell from about 51 cents for April 2025's RolmOCR to under 10 cents for October 2026's LightOnOCR-3-0.8B, a roughly fivefold drop in 18 months. At that rate, digitizing a million pages now costs $65 to $130 in GPU rental, sharply lowering the barrier to large-scale book and archive digitization.
DeepSeek has exposed the full capability stack behind its consumer app through its API, letting developers enable web search, file and PDF parsing, link reading, safety filtering and voice with a few parameters, or call standalone OCR, ASR and TTS endpoints. An open-source frontend demo with conversation-history import support means developers can closely recreate the entire DeepSeek App.
Nace AI has launched Drex, a document parser that uses decision models to compare each OCR block against the original page image and sends only mismatches back for correction. It scores 0.834 on the Parse Index, beating GPT-6 Astra's 0.813, at $7 per 1,000 pages versus $70.57.
Baidu has officially released and open-sourced the Wenxin Unlimited OCR 2 model, pushing document parsing limits to 'infinite' by processing over 3,600 pages in a single pass. The model natively supports layout segmentation and cross-page merging, achieving end-to-end SOTA performance on both MPDocBench and MarathonDocBench benchmarks.
The open-source model OXR (Optical Everything Recognition), with only 1.3B parameters, has achieved global first place in three major document parsing benchmarks including OmniDocBench and olmOCR-Bench. By incorporating Recursive Self-Improvement (RSI) mechanisms, OXR significantly enhances accuracy and speed for complex layouts, handwritten formulas, and chemical structures, processing 100 pages in just 101 seconds on a single RTX PRO 6000 GPU, thereby providing high-quality context extraction for agent infrastructure.
ModelBest has released and open-sourced MiniCPM-OCR 1.0, its first dedicated OCR model. With only 1.3 billion parameters, it surpasses major tech giants' models in both parsing capability and efficiency, securing the top global rankings. This release marks a significant breakthrough for small-parameter models in vertical document understanding, offering a highly cost-effective open-source alternative.
The open-source Chinese model OXR (1.3B parameters) has secured the top position in three major OCR benchmarks: OmniDocBench, olmOCR-Bench, and DocParsingBench. Leveraging RSI technology, it demonstrates superior parsing capabilities for complex documents like handwritten notes and chemical structures, achieving end-to-end processing of 100 pages in just 101 seconds on a single RTX PRO 6000 GPU.
Erwin Nieuwlaar won the $50K First Letters Prize from the Vesuvius Challenge for identifying readable text in the carbonized scroll PHerc. 343. This milestone leverages virtual unwrapping and ink detection AI, signaling growing momentum in scaling ancient document recovery beyond initial prototypes.
该教程展示了如何将 Unsiloed 的结构化文档提取能力与 TypeSafe Jev 的类型化问答引擎结合,以解决 OCR 识别正确但语义逻辑错误的痛点。Jev 通过定义状态(State)和具体问题(Questions),返回概率值或枚举选择等强类型答案,从而自动标记需人工复核的字段。文章提供了基于 Token 用量的成本估算示例,强调了在数据管道中引入独立验证层的重要性。