EVIE
modelYour notes
Evidence-Vector-Informed Embedding: Tencent's multi-vector, late-interaction visual document retrievers, built on ColQwen3.5-style backbones from Qwen3.5. EVIE-8B is the flagship teacher, producing 4,096-dimensional per-token embeddings with bidirectional full attention over page images and queries; EVIE-4.5B is the distilled student with Prefix-MRL and HAC for cheaper serving. Tencent's own leaderboard table places the pair at the top of ViDoRe V1, V2, and V3 among open multi-vector retrievers. Weights, inference pipelines, and evaluation suites are open under Apache 2.0.
The promised technical report landed on 2026-10-08 as arXiv 2610.11553, "EVIE: Evidence-Vector-Informed Embeddings for Visual Document Retrieval" (Zifei Wang, Wei Wen, Qiang Ji, Qian-Wen Zhang, Ruizhi Qiao, Xing Sun), and the GitHub README and both model cards now cite it.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| EVIE-8B | — | Flagship teacher; 4096-D per-token embeddings, bidirectional attention |
| EVIE-4.5B | — | Student with Prefix-MRL and HAC |