Tencent releases EVIE-8B and EVIE-4.5B visual document retrieval models

Tencent released two visual document retrieval models, EVIE-8B and EVIE-4.5B, on Hugging Face. The 8B model reports 66.75 nDCG@10 on the ViDoRe V3 benchmark, which the company describes as state-of-the-art retrieval accuracy, while the smaller 4.5B model scores 66.02 using a single-projection Prefix-MRL head.

The 8B model produces 4096-dimensional, per-token multi-vector embeddings meant to preserve fine-grained layout, typography, charts, and table structure. It also serves as a teacher, supplying capacity-aware relation and margin distillation targets for the lightweight 4.5B model. Both were validated across 138 tasks spanning ViDoRe V1, V2, V3, and JinaVDR, using four metric families: nDCG, Recall, MAP, and MRR.

The 4.5B model's Prefix-MRL design uses a single 2048-dimensional linear projection that can be truncated at runtime to any of 64, 128, 256, 512, 1024, or 2048 dimensions without separate models. A training-free Hierarchical Agglomerative Clustering index compresses token counts from roughly 750 down to 3, shrinking the retrieval index footprint.