UNISOUND (09678) launches the intelligent basic large model "Unisound U1-OCR", officially opening the OCR 3.0 era.

On February 26th, UNISOUND (09678) announced the official launch of the document intelligent foundational large model "Unisound U1-OCR". As the first industrial-grade document intelligent base, this model officially opens the OCR 3.0 era. Building on the understanding of layout, it further delves into the deep semantics of documents, achieves automatic classification and business-level information extraction, completes a qualitative leap from "character perception" to "document cognition", marking the transition of AI from simply "reading words" to "understanding business logic". Unisound U1-OCR is a document intelligent understanding model that reaches the international state-of-the-art level (SOTA), performing industry-leading in various authoritative tests. Its core advantage lies in breaking through the bottleneck of traditional models that "only read text, but don't understand layout", being able to "understand" complex documents like human experts. To meet the new requirements for document-level structured extraction in the OCR 3.0 era, Unisound U1-OCR adopts the ViT + LLM architecture, with the visual encoder part using the NaViT architecture, achieving dynamic document resolution processing. The model scale is at the 3 billion level, balancing model computational efficiency with the requirement for understanding deep semantic information in documents. The model introduces several innovative initiatives: it introduces the "semantic-driven + dynamic focus" strategy, automatically constructing a "semantic map" of documents, accurately identifying the hierarchical relationship between titles, graphs, and body text. It possesses the wisdom of "understanding structure first, then reading content"; it has a keen "spatial perception" and can actively understand the spatial layout between elements, combined with dynamic resolution technology to accurately restore the document structure. In addition, it uses Multi-Token Prediction (MTP) technology to consider the probability distribution of multiple future Tokens when predicting the current Token, significantly improving the logical coherence of long documents. Combined with a full-task reinforcement learning strategy, it enhances the model's global foresight of the layout structure and improves the model's generation efficiency by over 80% in the reasoning stage. At the business level, the model focuses on industrial-level scenario demands, building four core capabilities: precise traceability, business integration, secure and efficient deployment, and super strong adaptation. It truly meets the all-round needs of real enterprise business, moving from "understanding" to "execution" of business implementation. Unisound U1-OCR opens the OCR 3.0 era, not only a revolution in document intelligence but also a key step for UNISOUND towards AGI. The company will use multimodal documents as a knowledge entry point, endowing machines with autonomous reasoning and evidence tracing capabilities, advancing AI from perception to cognition. In the future, UNISOUND hopes to build a general intelligence that can read, think, and solve complex problems like humans, making every document a ladder to AGI.