Tencent Unveils EVIE Visual Document Retrieval Models

Tencent has released a new series of visual document retrieval models, designated as EVIE-8B and EVIE-4.5B, designed to improve the accuracy of retrieving information from complex visual documents such as charts, tables, and structured layouts.

Performance Benchmarks and Technical Specifications

According to technical documentation hosted on Hugging Face, the larger EVIE-8B model achieved a state-of-the-art (SOTA) performance score of 66.75 nDCG@10 on the ViDoRe V3 benchmark. The model utilizes high-capacity 4096D representations, employing full per-token multi-vector embeddings to preserve fine-grained details including typography and table structures.

The release includes a smaller, more efficient version, EVIE-4.5B, which utilizes a single-projection Prefix-MRL architecture. This model achieved a 66.02 score on the ViDoRe V3 benchmark. A notable feature of the 4.5B model is its “Prefix-MRL” elasticity, which allows for runtime truncation of the 2048D linear projection into various dimensions, including 64, 128, 256, 512, 1024, and 2048, without requiring separate model files.

To address storage concerns, the models utilize Hierarchical Agglomerative Clustering (HAC), a training-free method that compresses token counts from approximately 750 down to 32 vectors per page. This compression reduces index storage requirements to 3.81 GiB per million pages.

The models were validated across 138 tasks spanning four standard metric families: nDCG, Recall, MAP, and MRR u/1. The evaluation covered multiple benchmarks, including ViDoRe V1, V2, V3, and JinaVDR.

Distillation and Implementation

The EVIE-8B model serves as a “teacher” foundation for the EVIE-4.5B model. Through the EVIE-ARD distillation recipe, the larger model provides capacity-aware relation and margin distillation targets, allowing the lightweight student model to reproduce the performance characteristics of the 8B version.

While the models are designed to process natural language queries regarding visual content—such as asking for specific labels on an axis—technical discussions on platforms like Reddit have noted that the model outputs numeric values representing query-document relevance scores rather than direct text answers.

Broader Industry Context

The release of these models occurs as Chinese technology firms continue to expand their AI capabilities. Recent reports indicate that major players like Tencent and ByteDance have placed bulk orders for upcoming Huawei chips to support large-scale model training, including DeepSeek’s V4 model. This follows broader trends in the sector, including Huawei’s development of the Ascend-branded AI chip infrastructure and its recent launch of the CloudMatrix 384 server as an alternative to Nvidia’s AI stack.