Tencent released the EVIE-4.5B visual document search model

Company Tencent introduced a compact model EVIE-4.5B designed for visual document search tasks. The model has 4.61 billion parameters and is focused on high accuracy processing of graphical data in corporate and research environments.

In testing on the ViDoRe V3 benchmark, the architecture showed a result of 66.02 nDCG@10. The developers implemented a unique single-projection Prefix-MRL structure in the model, which allows for efficient matching of visual document features with user text queries.

The technical implementation includes HAC token compression technology, which optimizes the token compression process. This solution allows the model to maintain high performance when working with visually complex documents that require in-depth analysis of page structure and tabular data.

The model is already available for download in open access on the Hugging Face platform. The source code and model weights are distributed for integration into automated search and information extraction systems from unstructured visual sources.

The release of EVIE-4.5B expands the company's line of specialized solutions for working with visual data. The use of optimized token compression architectures reduces computational costs while maintaining search accuracy, which is critical for scalable document processing systems.