UNISOUND (09678) U1-OCR architecture paradigm upgrade Open standard API Refactoring OCR 3.0 era

On February 26th, UNISOUND (09678) released its first industrial-grade document intelligent base model, Unisound U1-OCR, ushering in the era of OCR 3.0 and laying a solid foundation for the iterative upgrading of the U1-OCR series models. Today (April 21st), after undergoing a bottom-up architecture reconstruction and extensive real-world scene polishing, the capabilities of UNISOUND U1-OCR have evolved again, leading to the launch of a series of models. At the same time, this model has been fully launched on the UNISOUND Token Hub large model service platform, offering standardized APIs and supporting easy integration and on-demand usage. It adopts a token billing model, significantly reducing the cost and deployment threshold for enterprise access, making the document intelligence capabilities of the OCR 3.0 era more accessible to a wider range of industries. The architecture paradigm of UNISOUND U1-OCR has been upgraded, abandoning traditional NMS and using a unified structure with refined solutions to solve cascading errors and achieve a qualitative leap in complex layout analysis. With its technical capabilities endorsed by authoritative certifications and multiple core papers selected for ACL2026, along with the top-ranked authoritative datasets, its performance can be verified and traced. Furthermore, it is adaptable to various industry scenarios, supporting the processing of complex documents in fields like finance, healthcare, education, and transportation, achieving structure understanding and sequential recovery in one go. A typical challenge in parsing complex documents is that structural information is not consistently organized and is difficult to efficiently deliver to downstream modules. The goal of U1-OCR is not just to "recognize text," but to effectively address the challenges of structural understanding and reading sequence recovery in complex document pages. Addressing the industry-wide problem, UNISOUND has employed a scenario-specific parsing design in U1-OCR, which can be fundamentally broken down into two core subtasks: structural recognition, which determines the content types of each area on a page and identifies the areas to be retained, and sequential reasoning, which plans a rational reading path for the retained areas. To address these tasks, U1-OCR was designed with specific key technologies, achieving leading results on multiple public authoritative datasets and providing a more stable and reliable processing method for the often overlooked detector-to-parser handoff in real business scenarios. Experimental results also show that in pages with more complex structures and layout variations, the U1-OCR model matrix can more efficiently handle issues such as boundary determination, category differentiation, and overall structure recovery, accurately achieving the design goal of stabilizing competitive candidate hypotheses as input structures for parsers. This also signifies that document parsing is transitioning from simple OCR text recognition to a more business-demand-sensitive document understanding capability. With the full launch of U1-OCR on the UNISOUND Token Hub large model service platform and the simultaneous opening of standardized APIs and one-click invocation functions, the threshold for using document intelligence technology will be further lowered, providing efficient and accurate document parsing services for industries such as healthcare, transportation, finance, and education, and assisting industries in smoothly achieving digital transformation and upgrading.