Tencent has released EVIE-8B, a high-capacity visual document retrieval model featuring 4096-dimensional token representations and bidirectional full attention. The model serves as a teacher foundation for the lightweight EVIE-4.5B Prefix-MRL variant and achieves state-of-the-art performance on visual document retrieval benchmarks.
- Achieves 66.75 nDCG@10 on ViDoRe V3 Public, delivering industry-leading accuracy.
- Provides full per-token multi-vector embeddings that preserve fine-grained layout, typography, charts, and table structures.
- Validated across 138 tasks in ViDoRe V1, V2, V3, and JinaVDR using four standard metric families.
- Trained on 775,635 document-query pairs with hard negatives mined and classified into answerable, ambiguous, and strict categories.
The release includes the model weights, inference engine, and evaluation harness under an Apache-2.0 license to facilitate high-fidelity visual document retrieval.