tencent/EVIE-Preview-4.5B
Visual Document Retrieval • 5B • Updated • 351 • 40
EVIE — Visual Document Retrieval
Multilingual visual document retrieval with native 128-dimensional multi-vector token embeddings. Rank #1 on ViDoRe V3 and on ViDoRe V1+V2.
Note 4.54B params on Qwen3.5-4B, ColBERT-style late interaction with 128D token vectors. ViDoRe V3 public 65.36 nDCG@10, ViDoRe V1+V2 85.77 nDCG@5. Trained at 768 visual tokens per page; the 1,792 tier is pure test-time extrapolation on the same weights.