Tencent has released EVIE-4.5B, a high-performance multilingual visual document retrieval model featuring dynamic Prefix-Multivector Retrieval Latency (Prefix-MRL) and training-free Hierarchical Agglomerative Clustering (HAC) token compression.
- EVIE-4.5B achieves 66.02 on ViDoRe V3 Public using a single-projection Prefix-MRL, while the larger EVIE-8B variant scores 66.75.
- The model supports runtime dimension truncation from 2048D down to 64D without requiring separate models, maintaining high retrieval accuracy across all sizes.
- Training-free HAC compresses token counts from approximately 750 vectors per page to 32, reducing index storage to 3.81 GiB per million pages.
- The model was evaluated across 138 multilingual tasks on ViDoRe V1, V2, V3, and JinaVDR, demonstrating strong performance in English, French, German, Spanish, Arabic, and Chinese.
- EVIE-4.5B is distilled from the EVIE-8B teacher using the EVIE-ARD recipe, which employs anchor-preserving, capacity-aware relation distillation.
The release provides a flexible solution for visual document retrieval that balances high precision with significantly reduced storage requirements through dynamic embedding dimensions and token compression.