Unisound has officially launched the Unisound U1-OCR foundation model for document intelligence. As the first industrial-grade foundation for document intelligence, the model breaks through the boundaries of conventional document processing and sets a new industry benchmark with five core strengths: SOTA performance, trustworthy and verifiable results, out-of-the-box usability, efficient deployment, and strong adaptability.

I. A Technological Leap: From OCR 2.0 to OCR 3.0

Document Intelligence refers to the use of AI technologies to automatically read and understand document images, including content recognition, comprehension, classification, and key information extraction. Traditional vision-based approaches (OCR 1.0, represented by CRNN) can only recognize text. The new generation of multimodal approaches (OCR 2.0, represented by VLMs) offers end-to-end layout understanding and text recognition. Unisound U1-OCR formally ushers in the OCR 3.0 era: beyond understanding document layouts, it identifies deeper semantics, performs automatic classification and business-level information extraction, and makes a qualitative leap from "character perception" to "document cognition."

II. Leading Performance: Global Top-Tier Results Across Multiple Authoritative Benchmarks

Unisound U1-OCR is a state-of-the-art (SOTA) document intelligence model. Its core strength is breaking through the conventional limitation of "reading text without understanding layout," enabling it to understand complex documents like a human expert. To meet OCR 3.0's new requirements for business-level structured document extraction, Unisound U1-OCR uses a ViT + LLM architecture. Its vision encoder employs NaViT to process documents at dynamic resolutions. At approximately 3B parameters, the model balances computational efficiency with deep semantic document understanding. It also introduces several innovations:

First, it has the intelligence to "understand structure before reading content." Conventional models often read rigidly in sequence, while Unisound U1-OCR pioneers a "semantic-driven + dynamic focus" strategy. Mirroring human reading habits, it first maps the hierarchy among tables of contents and headings, then extracts content as needed. The model automatically builds a "semantic map" of each document, accurately identifies the relationships among headings, charts, and body text, and extracts information clearly even from extremely disordered layouts.

Second, it has acute "spatial awareness." An enhanced spatial alignment module makes full use of the positional information of text on the page, allowing the model to actively understand the spatial layout among elements. Combined with dynamic resolution technology, it accurately reconstructs document structure across dense tables and mixed text-image layouts, eliminating the spatial confusion that caused previous models to associate content with the wrong elements.

The model also uses Multi-Token Prediction (MTP): when predicting the current token, it simultaneously considers probability distributions for multiple future tokens, substantially improving logical coherence in long documents. Combined with a full-task reinforcement learning strategy, MTP strengthens global foresight over layout structure and improves generation efficiency by more than 80% during inference. During training, a collaborative multi-task reinforcement scheme deeply aligns document structure reconstruction, document classification, and information extraction. The reinforcement strategy jointly optimizes semantics and coordinates, with targeted improvements to IoU accuracy in coordinate backtracking to suppress localization hallucinations and ensure the physical verifiability of outputs. Multi-level resolution perturbation and Mask sampling further improve the model's ability to understand document images across scenarioses.

Through these innovations, Unisound U1-OCR achieves industry-leading SOTA performance across multiple authoritative tests, delivering the leap from "recognizing text" to "understanding documents."

1. SOTA on OmniDocBench V1.5

On OmniDocBench V1.5, Unisound U1-OCR achieved a SOTA score of 95.1 (Figure 1), outperforming mainstream models including GLM-OCR, DeepSeek-OCR2, Gemini 3 Pro, and GPT-5.2 and setting a new benchmark for both accuracy and generalization.

Figure 1. OmniDocBench V1.5 Score Comparison for Unisound U1-OCR

2. SOTA on D4LA

On D4LA, the model achieved an F1 score of 90.8 (Figure 2), substantially outperforming DocLayout-YOLO (87.3) and PP-StructureV3 (86.0). Without fine-tuning, it accurately parses 11 types of highly complex documents, including academic papers and financial statements.

Figure 2. Cross-Model Comparison of Document Layout Parsing on D4LA

3. SOTA on DocLayNet

On DocLayNet, the model achieved an F1 score of 95.9 (Figure 3), surpassing MinerU 2.5, PP-StructureV3, and other models. It shows clear advantages and exceptional robustness on challenging tasks such as table recognition, cross-page association, and small-text detection.

Figure 3. Cross-Model Comparison of Document Layout Parsing on DocLayNet

4. SOTA on Business-Relevant evalsuations

In internal business tests (Figure 4), its information extraction and document classification capabilities surpassed mainstream general-purpose commercial and open-source models including Gemini 3 Flash and Qwen-235B-VL. Its lead was especially pronounced in demanding business scenarioses such as medical admission records and discharge summaries. With just 3B parameters, Unisound U1-OCR outperformed much larger general-purpose VLMs. Compared with smaller document-parsing task models, its innovations also deliver stronger deep semantic understanding for capabilities such as business-level information extraction.

Figure 4. Cross-Model Comparison of Document Classification and Information Extraction on Business Datasets

III. Built for Real-World Scenarioses: Four Core Capabilities Take U1-OCR from "Understanding" to "Execution"

As a foundational document intelligence model ushering in the OCR 3.0 era, Unisound U1-OCR not only achieves SOTA results on multiple general benchmarks but also addresses the needs of industrial-grade scenarioses through four core capabilities, enabling business deployment from "understanding" to "execution."

1. Trustworthy and Traceable: Precise Grounding, Verifiable Results

The model introduces a coordinate-text-semantics fusion architecture that enables pixel-level precise grounding and complete evidence chains. While extracting information, the system accurately identifies its source locations in the document, making result review fully transparent and traceable. This technically ensures the trustworthiness of document-processing results and resolves the industry's longstanding problem of unverifiable outputs.

For example, in enterprise review scenarioses, reviewers no longer need to search through the original document line by line. Clicking an extracted result highlights its source locations in real time. This closed-loop human-AI collaboration reduces review time to seconds, minimizes human omissions, and delivers truly trustworthy AI.

2. Business Integration: Out of the Box and Agent Ready

General-purpose OCR tools have limitations in specialized domains. For example, understanding the logical relationship among "out-of-pocket payment category 1," "out-of-pocket payment category 2," and "fully self-paid" on medical insurance settlement statements, or validating amounts written in numerals and words in contracts, requires domain knowledge.

Built on the foundation model, Unisound U1-OCR incorporates Unisound's accumulated industry knowledge in healthcare, finance, and other fields, enabling multi-field cross-validation based on business logic. In internal business tests, classification accuracy exceeded 99% across more than 50 common business document types.

3. Efficient Deployment, Secure and Controllable

The model deeply supports private and offline deployment and operates reliably without external network access, making it well suited to the data privacy requirements of high-security sectors such as government, healthcare, and finance. Optimizations including layout-level parallel decoding and Multi-Token Prediction allow a document of more than ten pages to be processed in seconds. This efficient processing brings industrial-grade document intelligence within easy reach.

4. Exceptional Adaptability for Complex Scenarioses

Across extremely complex real-world enterprise scenarioses, including nonstandard photos, bent or blurred documents, intricate layouts, and mixed-language content, Unisound U1-OCR maintains stable, high-accuracy performance. It eliminates conventional technologies' dependence on standardized documents and adapts to the full range of real business needs.

IV. Validated in Real-World Scenarioses

1. Trusted Information Extraction Cases

Case 1: The model deeply understands business requirements and makes extracted data immediately usable. When extracting medical expense statements, it automatically understands semantics and accommodates wording differences among hospitals. For example, it maps labels such as "Total" and "Total Amount" in source images to the database field "Total Expenses." Following business-field extraction rules, it accurately removes irrelevant high-level categories so results can be written directly to the database. The model also supports pixel-level coordinate grounding, linking each extracted result to its source position in the image with different colors, as shown. This transparent and trustworthy system upgrades traditional full-document rereading into point-specific confirmation within seconds, improving business efficiency while safeguarding the accuracy of stored data.

2. Business Knowledge Integration Cases

Case 2: The model automatically recognizes and segments multiple documents within a single image, accurately distinguishing mixed files such as medical records and prescriptions. Without manual pre-classification, mixed documents can be automatically archived and extracted with one click, making large volumes of disorganized documents simple and efficient to process.

Case 3: Even with suboptimal documents affected by photographic obstruction or missing content, the model uses deep semantic insight to classify them accurately. Its strong compatibility with complex long-tail scenarioses eliminates blind spots in real-world operation and substantially improves automated processing success rates.

3. Complex Layout Recognition Cases

Case 4: Conventional approaches often lose the reading order when parsing complex multi-column layouts in newspapers and journals, where text and images are interwoven. Unisound U1-OCR breaks through this limitation: instead of mechanically scanning pages in a fixed sequence, it combines contextual semantics with layout logic as a human reader would, automatically determines how paragraphs connect, and reconstructs a content flow that matches natural reading habits.

Case 5: For pages covered with watermarks or set at an angle, the model automatically performs "image purification," intelligently removing watermark interference and precisely correcting distorted layouts. No matter how cluttered the source image is, it outputs a standardized document with an orderly layout and clear content, providing a clean foundation for subsequent recognition and eliminating interference risks.

Case 6: When faced with complex tables spanning rows and columns or containing nested structures, the model not only parses content accurately but also fully preserves the original row-column structure and logical relationships. Results are immediately usable without secondary adjustment, whether the source is a statistical report or an engineering drawing.

Unisound U1-OCR ushers in the OCR 3.0 era, marking AI's leap from simply "recognizing characters" to "understanding business logic." This is not only an innovation in document intelligence but also a critical step in Unisound's journey toward AGI. With multimodal documents as a gateway to knowledge, we will give machines autonomous reasoning and evidence-tracing capabilities, advancing AI from perception to cognition. Looking ahead, Unisound aims to build general-purpose AI agents that can read, think, and solve complex problems like humans, making every document an intelligent step toward AGI.