Tencent released the EVIE-8B visual document search model

Company Tencent introduced a specialized model EVIE-8B designed for searching and extracting information from visual documents. The model has 8.41 billion parameters and demonstrates advanced performance metrics in its category.

In testing on the ViDoRe V3 benchmark, the architecture showed a result of 66.75 nDCG@10. The technical implementation of the system is based on the ColQwen3.5 architecture, which provides high accuracy in processing visual data. The model uses 4096D embeddings for each token, allowing for efficient indexing of complex document structures.

The development is focused on visual search tasks (visual document retrieval), where high accuracy of matching the query with the graphical representation of the content is required. The model is already available for use and integration into specialized search systems.

The release of EVIE-8B expands the toolkit for working with unstructured visual data. The use of the ColQwen3.5 architecture makes it possible to achieve a balance between the size of the model and the quality of the search, which is critical for high-load document processing systems.