ColNanoVDR Distills Multi-Vector Doc Retrieval Without Encoding Documents

ColNanoVDR introduces document-free distillation to multi-vector visual document retrieval, training a compact 149M model that preserves 95% retrieval accuracy while encoding 26x faster.

2026-09-29 ~ 2026-09-29 · 2 related posts

• ColNanoVDR distills multi-vector visual document retrieval without documents, keeping 95% NDCG@5 at 149M params — nanovdr · 2026-09-29

1 near-duplicate retellings: _reachsumit